Loading video...

Video Failed to Load

Go Home

LOOP RAT: THE TECH UNDER AN AUTONOMOUS CODING LOOP no framework, no service, no database a dozen files of bash and python, plus one per loop one shift runs six phases: > preflight - won't start if state/HALT exists, spend is gone, or the loop's already running > act...

polydao's profile picture

Mr. Buzzoni

23,020 subscribers

11,827 views • 10 days ago •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

KIMI K3 + OBSIDIAN + LOOP ENGINEERING = A VAULT THAT RUNS ITSELF the core idea: the vault is the loop's state, not the chat window everything Kimi K3 knows lives in a .md file the loop: > capture - a thought lands in 00-inbox > context - K3 pulls links, tags, and neighbouring notes > draft - edits happen inside a git worktree, never the live vault > review - a critic agent checks the diff before anything ships > commit - appended to the vault, nothing gets rewritten the key insight: frontmatter fields like supports, contradicts, and supersedes are graph edges, not metadata - the note format is the write API start with a plain loop, it runs about 2-4x the cost of one direct call > only move to a full graph once state has to outlive the session, several agents need to coordinate, or you have to explain what changed - that jump can run 10-50x one review assistant climbed from 55% to 72% to 84% just by moving through these shapes in order
0:20

KIMI K3 + OBSIDIAN + LOOP ENGINEERING = A VAULT THAT RUNS ITSELF the core idea: the vault is the loop's state, not the chat window everything Kimi K3 knows lives in a .md file the loop: > capture - a thought lands in 00-inbox > context - K3 pulls links, tags, and neighbouring notes > draft - edits happen inside a git worktree, never the live vault > review - a critic agent checks the diff before anything ships > commit - appended to the vault, nothing gets rewritten the key insight: frontmatter fields like supports, contradicts, and supersedes are graph edges, not metadata - the note format is the write API start with a plain loop, it runs about 2-4x the cost of one direct call > only move to a full graph once state has to outlive the session, several agents need to coordinate, or you have to explain what changed - that jump can run 10-50x one review assistant climbed from 55% to 72% to 84% just by moving through these shapes in order

Mr. Buzzoni

10,680 views • 24 days ago

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.
0:18

Don't train the model, evolve the harness. I read a brilliant blog post from Hugging Face where they took a frozen open model scoring 0% on a hard legal agent benchmark, left its weights alone, and let an automated loop rewrite only the code around it. That code layer is the harness, the runtime wrapper that feeds the model context, runs its tool calls, and decides when a run ends. By the time the loop finished, the system had essentially matched Sonnet 4.6 on the benchmark's headline metric, at roughly 7x lower cost per task. Zero weights changed. The gain existed because of where the model was failing. The judge only grades files saved in the right place under the exact requested filename, and the model kept doing the legal analysis correctly, then saving it under the wrong name, dropping it in a scratch folder, or never writing it at all. So the 0% was never measuring legal reasoning. It was measuring the harness. Hand-tuning that layer is slow and model-specific, so they automated it. A Claude proposer adds exactly one mechanism per iteration, and an outer loop keeps it only if it clearly beats the current best, so accepted mechanisms compound. What the loop discovered says a lot about where agents actually fail. → The biggest single gain was file handling, not intelligence. An automatic step that lands the deliverable exactly where the judge expects it beat every prompt change, with zero extra model tokens. → Code fixes transferred across models, prompt playbooks did not. The same harness lifted a smaller model from the same family by 14 points, but the tuned prompts hurt a different model family on tasks it could already finish. → The harness mattered more than anything else. Same model, same judge, same tasks, and five different harnesses scored anywhere between 3.5% and 80.1%. The gains do eventually flatten, and the remaining misses look like real capability gaps. At some point the wrapper runs out of tricks and the model has to carry the work. But the lesson holds. A benchmark score measures the model and its harness together, and until the harness is fixed, it's impossible to know which one failed. I highly recommend reading this: I also wrote a deep dive on agent harness engineering a while back, covering the orchestration loop, tools, memory, context management, and everything that turns a stateless LLM into a capable agent. The article is quoted below.

Akshay 🚀

244,990 views • 2 months ago

sorry, they just did WHAT someone gave a machine one disease name, the leading cause of blindness in the developed world with 1.5 million americans already in its path, and it came back pointing at a drug that has sat in pharmacies for years under a different label: 551 papers read in 30 minutes against the 294 hours a human would have needed, and the loop that did it is public on GitHub most agent setups answer one question at a time, so the ceiling on the work is the quality of the question you happened to think of this one was handed a single question and wrote the second one itself. turns out that follow-up is where the real find was: a target called ABCA1, upregulated threefold, in an experiment no human ordered i read the whole paper looking for the trick, and the trick is structural. that is the second question, and it is the gap between an assistant and a factory: - hand the loop a field rather than a task: it was given a disease, and choosing the mechanism was part of its job - make it rank before it spends: 151 papers in, ten candidate mechanisms out, scored against each other before anything touched a bench - split reading from judging, so the agent that forms the theory is a different agent from the one grading it - close every cycle on physical reality: the verdict was an experiment, and another model's opinion was never allowed to stand in for one - feed each result back as the next question rather than a log line, which is the step almost nobody builds - search what already passed inspection first: the winner was an approved compound with a safety file already on record - write down what the round learned before opening the next one, so round two starts where round one stopped my read, and i think it is the uncomfortable one: reading was the entire bottleneck in that field, and everybody spent the decade optimising the writing. people ran every physical experiment here, the analysis agent needs a domain expert writing its prompts, and the authors decline to call this the leap it resembles. the thinking got replaced, and the hands did not so the question i cannot answer for my own setup: which step of your loop still stops dead until you sit down and type something bookmark this one. the four parts that turn one model into a line that runs like this, the queue, the rooms, the write permissions and the gate, are built file by file in the piece below ↓
0:15

sorry, they just did WHAT someone gave a machine one disease name, the leading cause of blindness in the developed world with 1.5 million americans already in its path, and it came back pointing at a drug that has sat in pharmacies for years under a different label: 551 papers read in 30 minutes against the 294 hours a human would have needed, and the loop that did it is public on GitHub most agent setups answer one question at a time, so the ceiling on the work is the quality of the question you happened to think of this one was handed a single question and wrote the second one itself. turns out that follow-up is where the real find was: a target called ABCA1, upregulated threefold, in an experiment no human ordered i read the whole paper looking for the trick, and the trick is structural. that is the second question, and it is the gap between an assistant and a factory: - hand the loop a field rather than a task: it was given a disease, and choosing the mechanism was part of its job - make it rank before it spends: 151 papers in, ten candidate mechanisms out, scored against each other before anything touched a bench - split reading from judging, so the agent that forms the theory is a different agent from the one grading it - close every cycle on physical reality: the verdict was an experiment, and another model's opinion was never allowed to stand in for one - feed each result back as the next question rather than a log line, which is the step almost nobody builds - search what already passed inspection first: the winner was an approved compound with a safety file already on record - write down what the round learned before opening the next one, so round two starts where round one stopped my read, and i think it is the uncomfortable one: reading was the entire bottleneck in that field, and everybody spent the decade optimising the writing. people ran every physical experiment here, the analysis agent needs a domain expert writing its prompts, and the authors decline to call this the leap it resembles. the thinking got replaced, and the hands did not so the question i cannot answer for my own setup: which step of your loop still stops dead until you sit down and type something bookmark this one. the four parts that turn one model into a line that runs like this, the queue, the rooms, the write permissions and the gate, are built file by file in the piece below ↓

Argona

32,475 views • 1 month ago

FIVE LAYERS OF AGENT ENGINEERING, EACH ONE WRAPS THE ONE BELOW IT. IF YOU SKIP LAYER 2, YOUR LAYER 5 WILL LOOK BROKEN WHEN IT IS ACTUALLY JUST STANDING ON NOTHING. for weeks i debated harness vs loop vs graph like they were competing choices. then a stack diagram made the shape obvious. they are not choices. they are floors. 01 | prompt engineering. the message. unit of work: one input. inputs are role, instructions, examples, format. output is a single raw response. 02 | context engineering. the memory. unit of work: what stays in the window. a curator selects, compresses, and drops from query, docs, memory, prior turns, and tool outputs before the prompt runs. 03 | harness engineering. the machine. unit of work: the machine itself. gather (context + prompt) → LLM → tools or sub-agents → verifier → final response. the article calls this the operating environment. 04 | loop engineering. the system. unit of work: the run. goal + success criteria + max iterations + budget + completion check wrap around one harness pass. failed pass appends results to context and retries. 05 | graph engineering. the topology. unit of work: the graph run. goal + nodes + edges + state schema. graph routes to agent nodes, tool nodes, or human approval. a reviewer node with a different model and fresh context checks the final answer. the wrapping is the whole point. layer 5 assumes layer 4 works. layer 4 assumes layer 3 works. skip layer 2 and layer 3's verifier keeps failing without a clear reason. this is why swapping the model is a one-day project and swapping the stack is a quarter. the model is the commodity. the five layers around it are the engineering. full three-layer breakdown of the top of the stack (harness, loop, graph) in the post below.
0:20

FIVE LAYERS OF AGENT ENGINEERING, EACH ONE WRAPS THE ONE BELOW IT. IF YOU SKIP LAYER 2, YOUR LAYER 5 WILL LOOK BROKEN WHEN IT IS ACTUALLY JUST STANDING ON NOTHING. for weeks i debated harness vs loop vs graph like they were competing choices. then a stack diagram made the shape obvious. they are not choices. they are floors. 01 | prompt engineering. the message. unit of work: one input. inputs are role, instructions, examples, format. output is a single raw response. 02 | context engineering. the memory. unit of work: what stays in the window. a curator selects, compresses, and drops from query, docs, memory, prior turns, and tool outputs before the prompt runs. 03 | harness engineering. the machine. unit of work: the machine itself. gather (context + prompt) → LLM → tools or sub-agents → verifier → final response. the article calls this the operating environment. 04 | loop engineering. the system. unit of work: the run. goal + success criteria + max iterations + budget + completion check wrap around one harness pass. failed pass appends results to context and retries. 05 | graph engineering. the topology. unit of work: the graph run. goal + nodes + edges + state schema. graph routes to agent nodes, tool nodes, or human approval. a reviewer node with a different model and fresh context checks the final answer. the wrapping is the whole point. layer 5 assumes layer 4 works. layer 4 assumes layer 3 works. skip layer 2 and layer 3's verifier keeps failing without a clear reason. this is why swapping the model is a one-day project and swapping the stack is a quarter. the model is the commodity. the five layers around it are the engineering. full three-layer breakdown of the top of the stack (harness, loop, graph) in the post below.

kocer

30,675 views • 14 days ago

this is worth more than most five figure courses 16 claude agents audit an entire repo at once, a second fleet re-checks every finding on fresh context, and the whole thing runs off one diagram instead of a prompt i ran it against my own code and got back 11 endpoints where i never checked who was logged in, 3 of which the verifier threw out before they ever reached me this is Graph Engineering, the layer above prompting, and it runs on the agent you already pay for: - write your plan out, then ask one question at every "and then": does the next step actually read what the previous one produced - the seams that fail that question were never dependencies, so those jobs run at the same time - the arrows that survive are your real edges, and the longest chain of them is your floor that no number of agents shortens - want it faster, cut a false edge instead of adding a worker - fan the independent work out, one agent per item, no shared state between them - send every finding to a separate agent on fresh context, because a model recognises its own writing 73.5% of the time and grades it kinder once it does - make that verifier check a real signal like a passing test, never the worker's own word that it finished - shard the fleet across worktrees so parallel workers stop overwriting each other, one rule frozen into every worker: never git stash, never git reset - merge only what came back verified, into one report instead of twenty open chats the catch is the ceiling. at 95% independent work 16 agents return 9.14x rather than the 16 you would guess, and even 256 only reach 18.6x, because the merge and the verify stay serial however wide you fan coordination itself is free plain code and every agent underneath it is billed, so start at twenty files and widen once it works bookmark this, the whole method with all six ready-to-run graphs is written out in the article ↓
0:15

this is worth more than most five figure courses 16 claude agents audit an entire repo at once, a second fleet re-checks every finding on fresh context, and the whole thing runs off one diagram instead of a prompt i ran it against my own code and got back 11 endpoints where i never checked who was logged in, 3 of which the verifier threw out before they ever reached me this is Graph Engineering, the layer above prompting, and it runs on the agent you already pay for: - write your plan out, then ask one question at every "and then": does the next step actually read what the previous one produced - the seams that fail that question were never dependencies, so those jobs run at the same time - the arrows that survive are your real edges, and the longest chain of them is your floor that no number of agents shortens - want it faster, cut a false edge instead of adding a worker - fan the independent work out, one agent per item, no shared state between them - send every finding to a separate agent on fresh context, because a model recognises its own writing 73.5% of the time and grades it kinder once it does - make that verifier check a real signal like a passing test, never the worker's own word that it finished - shard the fleet across worktrees so parallel workers stop overwriting each other, one rule frozen into every worker: never git stash, never git reset - merge only what came back verified, into one report instead of twenty open chats the catch is the ceiling. at 95% independent work 16 agents return 9.14x rather than the 16 you would guess, and even 256 only reach 18.6x, because the merge and the verify stay serial however wide you fan coordination itself is free plain code and every agent underneath it is billed, so start at twenty files and widen once it works bookmark this, the whole method with all six ready-to-run graphs is written out in the article ↓

Argona

157,312 views • 1 month ago

ANTHROPIC ENGINEER JUST SHOWS EXACTLY WHAT GRAPH ENGINEERING LOOKS LIKE WHEN A TASK RUNS THROUGH IT most people arguing about graphs online have never actually watched one execute Task → Researcher → Planner → Writer + Code Agent → Reviewer → Deploy six nodes, one shared state, graph completes itself while you watch loop mode gets disabled the second the task splits into real specialties, one agent stops trying to do everything at once reviewer catches a failure, kicks it straight back to the exact node that broke - not a full restart, no lost context the graph is not a fancier loop - it is the org chart your agents were missing bookmark this and watch it run, then read the article below to see why the timeline just found what production systems already do
0:10

ANTHROPIC ENGINEER JUST SHOWS EXACTLY WHAT GRAPH ENGINEERING LOOKS LIKE WHEN A TASK RUNS THROUGH IT most people arguing about graphs online have never actually watched one execute Task → Researcher → Planner → Writer + Code Agent → Reviewer → Deploy six nodes, one shared state, graph completes itself while you watch loop mode gets disabled the second the task splits into real specialties, one agent stops trying to do everything at once reviewer catches a failure, kicks it straight back to the exact node that broke - not a full restart, no lost context the graph is not a fancier loop - it is the org chart your agents were missing bookmark this and watch it run, then read the article below to see why the timeline just found what production systems already do

leopardracer

53,328 views • 1 month ago

I STOPPED REVIEWING MY OWN AGENT, SOMETHING ELSE DOES IT NOW I used to read every diff it produced and approve most of them, because an agent grading itself always says the work is good. -> Now a second model with different instructions tries to break the work first, and I only read what survived. Here is what is actually in the folder that took over the night shift: • the brief > CONTRACT.md -- what it may touch, and what it may never touch. > VISION.md -- the destination, so turn 47 still knows why it started. • the gate > judge/ -- a different model, never the one that wrote the code. > break-it.md -- it opens the page, clicks, screenshots, reports back. > -- no opinion, just zero or non-zero. > shift.yml -- 03:30 every night, laptop closed. • the memory > receipts/ -- one folder per night, dated and graded. > STATE.md -- where it stopped and what it escalated. > lessons.log -- the flaky test, written down once instead of rediscovered weekly. • the brakes > caps.json -- turn limit, retry limit, spend limit. > -- written on day one, used never. The generator decides what your loop can produce -> The judge decides what it refuses to produce. One of those is the part everyone builds -> The other is why most loops quietly fail. Bookmark it & Read Full breakdown below ↓
0:18

I STOPPED REVIEWING MY OWN AGENT, SOMETHING ELSE DOES IT NOW I used to read every diff it produced and approve most of them, because an agent grading itself always says the work is good. -> Now a second model with different instructions tries to break the work first, and I only read what survived. Here is what is actually in the folder that took over the night shift: • the brief > CONTRACT.md -- what it may touch, and what it may never touch. > VISION.md -- the destination, so turn 47 still knows why it started. • the gate > judge/ -- a different model, never the one that wrote the code. > break-it.md -- it opens the page, clicks, screenshots, reports back. > -- no opinion, just zero or non-zero. > shift.yml -- 03:30 every night, laptop closed. • the memory > receipts/ -- one folder per night, dated and graded. > STATE.md -- where it stopped and what it escalated. > lessons.log -- the flaky test, written down once instead of rediscovered weekly. • the brakes > caps.json -- turn limit, retry limit, spend limit. > -- written on day one, used never. The generator decides what your loop can produce -> The judge decides what it refuses to produce. One of those is the part everyone builds -> The other is why most loops quietly fail. Bookmark it & Read Full breakdown below ↓

slash1s

37,556 views • 1 month ago

This Chinese developer launched Llama 70B locally on a MacBook on a plane and for a full 11 hours without internet ran client projects. He was sitting by the window on a transatlantic flight with a MacBook Pro M4 with 64 GB of memory. WiFi on board cost $25 for the flight. He declined. No cloud API, no connection to Anthropic or OpenAI servers, no internet at all. Just a local Llama 3.3 70B on bf16 and his own orchestrator script. The model runs through llama.cpp. Generation speed, 71 tokens per second. Context around 60,000 tokens. Memory usage, 48.6 GiB out of 64. Battery at takeoff, 3 hours 21 minutes. And he gave the orchestrator this system prompt before takeoff: "You are an offline orchestrator running on a single MacBook. There is no network. The only resources you have are local files in /Users/dev/work, the Llama 70B inference server at localhost:8080, and a battery budget of 3 hours 21 minutes. Process the queue at /Users/dev/work/queue.jsonl (one client task per line). For each task: draft → run local evals → save artefact to /Users/dev/work/done/. Save context checkpoints every 12 tasks so you can resume after a battery swap. Stop only on empty queue or when battery drops below 5%." So the system knows exactly what resources it is running on. It knows it has no connection to the outside world for the next 11 hours. It knows it has finite memory and a finite battery. It knows the human will not intervene until the plane lands. The system runs in 1 loop. Takes a task from the queue, runs it through inference, saves the artifact, writes a checkpoint. Task after task, just like that. And only when the battery drops below 5% does the orchestrator automatically pause, waits for the laptop to switch to the backup power bank, and continues from the last checkpoint. Here is what the system actually writes in his log during the flight: "saved context checkpoint 8 of 12 (pos_min = 488, pos_max = 50118, size = 62.813 MiB)" "restored context checkpoint (pos_min = 488, pos_max = 50118)" "prompt processing progress: n_tokens = 50 / 60 818" "task 37016 done | tps = 71 s tokens text → /Users/dev/work/done/proposal_westside.md" Outside the window, clouds, blue sky, and no WiFi. On the tray, 1 MacBook, an open terminal on 2 screens, and an inference server on localhost. From what I have observed, this is the cleanest offline AI workflow I have seen in the past year: 11 hours of flight, $0 for WiFi, and the entire client queue closed before landing.
0:10

This Chinese developer launched Llama 70B locally on a MacBook on a plane and for a full 11 hours without internet ran client projects. He was sitting by the window on a transatlantic flight with a MacBook Pro M4 with 64 GB of memory. WiFi on board cost $25 for the flight. He declined. No cloud API, no connection to Anthropic or OpenAI servers, no internet at all. Just a local Llama 3.3 70B on bf16 and his own orchestrator script. The model runs through llama.cpp. Generation speed, 71 tokens per second. Context around 60,000 tokens. Memory usage, 48.6 GiB out of 64. Battery at takeoff, 3 hours 21 minutes. And he gave the orchestrator this system prompt before takeoff: "You are an offline orchestrator running on a single MacBook. There is no network. The only resources you have are local files in /Users/dev/work, the Llama 70B inference server at localhost:8080, and a battery budget of 3 hours 21 minutes. Process the queue at /Users/dev/work/queue.jsonl (one client task per line). For each task: draft → run local evals → save artefact to /Users/dev/work/done/. Save context checkpoints every 12 tasks so you can resume after a battery swap. Stop only on empty queue or when battery drops below 5%." So the system knows exactly what resources it is running on. It knows it has no connection to the outside world for the next 11 hours. It knows it has finite memory and a finite battery. It knows the human will not intervene until the plane lands. The system runs in 1 loop. Takes a task from the queue, runs it through inference, saves the artifact, writes a checkpoint. Task after task, just like that. And only when the battery drops below 5% does the orchestrator automatically pause, waits for the laptop to switch to the backup power bank, and continues from the last checkpoint. Here is what the system actually writes in his log during the flight: "saved context checkpoint 8 of 12 (pos_min = 488, pos_max = 50118, size = 62.813 MiB)" "restored context checkpoint (pos_min = 488, pos_max = 50118)" "prompt processing progress: n_tokens = 50 / 60 818" "task 37016 done | tps = 71 s tokens text → /Users/dev/work/done/proposal_westside.md" Outside the window, clouds, blue sky, and no WiFi. On the tray, 1 MacBook, an open terminal on 2 screens, and an inference server on localhost. From what I have observed, this is the cleanest offline AI workflow I have seen in the past year: 11 hours of flight, $0 for WiFi, and the entire client queue closed before landing.

Blaze

1,842,664 views • 4 months ago

New open-source agent harness just landed! I got early access to TrueForge by TrueFoundry and have been running it locally for the past few days. The harness layer deserves as much attention as the model, and open source matters here because you can inspect the loop, run it on your own infrastructure, and swap to the latest or cheaper models. TrueForge handles the runtime work that makes an agent reliable. It drives the tool-calling loop, manages context, coordinates subagents, and executes code in a sandbox, with any model you choose. Every tool call re-sends the growing context to the model, so in practice the harness controls most of what an agent costs to run. A few things stood out from my testing and their published benchmarks. Vendor-Neutral by design. It runs OpenAI, Anthropic, and Google models alongside open-weight models like Kimi, GLM, and DeepSeek. Model routing is a setting, and you can send each task to the model that fits it. On a 14-task enterprise agent benchmark, it matched the accuracy of Claude Managed Agents running the same Opus 4.8 model at roughly 30% lower cost per run (3.8M tokens vs 10M for the same answers). Routing the same tasks to GLM-5.2 held accuracy and brought cost down by about 75%, around $3 per run instead of $12. Fully self-hosted and Open Source (MIT License). I had it running locally with one command, with sandboxed code execution working out of the box. It's time to own your agent harness. Thanks to TrueFoundry for partnering on this post.
0:16

New open-source agent harness just landed! I got early access to TrueForge by TrueFoundry and have been running it locally for the past few days. The harness layer deserves as much attention as the model, and open source matters here because you can inspect the loop, run it on your own infrastructure, and swap to the latest or cheaper models. TrueForge handles the runtime work that makes an agent reliable. It drives the tool-calling loop, manages context, coordinates subagents, and executes code in a sandbox, with any model you choose. Every tool call re-sends the growing context to the model, so in practice the harness controls most of what an agent costs to run. A few things stood out from my testing and their published benchmarks. Vendor-Neutral by design. It runs OpenAI, Anthropic, and Google models alongside open-weight models like Kimi, GLM, and DeepSeek. Model routing is a setting, and you can send each task to the model that fits it. On a 14-task enterprise agent benchmark, it matched the accuracy of Claude Managed Agents running the same Opus 4.8 model at roughly 30% lower cost per run (3.8M tokens vs 10M for the same answers). Routing the same tasks to GLM-5.2 held accuracy and brought cost down by about 75%, around $3 per run instead of $12. Fully self-hosted and Open Source (MIT License). I had it running locally with one command, with sandboxed code execution working out of the box. It's time to own your agent harness. Thanks to TrueFoundry for partnering on this post.

elvis

11,303 views • 25 days ago

girl humanoids become the hands of your autonomous agents ngl “imagine if your agents ran a girl humanoid” is not sci-fi bait anymore it is the missing physical layer under every todo list that still dies in a chat window → what the stack actually is your autonomous agents already research compare book draft and click they live in tabs calendars inboxes and apis what they never had was meters a soft body that can walk the kitchen lift the bag set the product in frame and finish the last ugly centimeter girl humanoids become that layer cameras for eyes tendon hands for grasp a dock for overnight charge and a policy link so the agent’s plan becomes motion instead of another notification you ignore you do not “chat with the robot” as the main product you assign the agent the girl is the effector teleop stays for the weird fail the brief stays yours → what that unlocks under your tasks home ops as an agent job check inventory open delivery pick the brand meet the courier put it away the agent owns the logic she owns the walk ugc and brand days the agent writes the shot list watches her cameras calls the next pour you approve the take like a creative director not a joystick babysitter admin with legs forms research returns scheduling on the laptop side then a body that can grab the box tape the label and leave it by the door multi-agent theater that finally touches reality one agent shops one edits the caption one runs qa on the take one soft girl executes the physical beat they all needed → where this leads the house becomes an api with a face staff is no longer only humans on payroll or silent appliances it is agents with rented hands in a knit suit clicking forever gets embarrassing if the plan holds the body finishes it teleop becomes exception handling the dock becomes the real producer credit also the hard edge who is allowed to drive her what she can buy film unlock and say battery still ends the shift precision still breaks on the tiny cruel tasks a frontier agent plus a home body needs a short leash or your kitchen becomes an unsupervised checkout spree → my take girl humanoids become the hands of your autonomous agents which means the win is not a cuter gait it is closing the loop from intent → plan → meters → done imagine is over the punchline is the dock assign the agent let the soft hands finish the room so back
0:10

girl humanoids become the hands of your autonomous agents ngl “imagine if your agents ran a girl humanoid” is not sci-fi bait anymore it is the missing physical layer under every todo list that still dies in a chat window → what the stack actually is your autonomous agents already research compare book draft and click they live in tabs calendars inboxes and apis what they never had was meters a soft body that can walk the kitchen lift the bag set the product in frame and finish the last ugly centimeter girl humanoids become that layer cameras for eyes tendon hands for grasp a dock for overnight charge and a policy link so the agent’s plan becomes motion instead of another notification you ignore you do not “chat with the robot” as the main product you assign the agent the girl is the effector teleop stays for the weird fail the brief stays yours → what that unlocks under your tasks home ops as an agent job check inventory open delivery pick the brand meet the courier put it away the agent owns the logic she owns the walk ugc and brand days the agent writes the shot list watches her cameras calls the next pour you approve the take like a creative director not a joystick babysitter admin with legs forms research returns scheduling on the laptop side then a body that can grab the box tape the label and leave it by the door multi-agent theater that finally touches reality one agent shops one edits the caption one runs qa on the take one soft girl executes the physical beat they all needed → where this leads the house becomes an api with a face staff is no longer only humans on payroll or silent appliances it is agents with rented hands in a knit suit clicking forever gets embarrassing if the plan holds the body finishes it teleop becomes exception handling the dock becomes the real producer credit also the hard edge who is allowed to drive her what she can buy film unlock and say battery still ends the shift precision still breaks on the tiny cruel tasks a frontier agent plus a home body needs a short leash or your kitchen becomes an unsupervised checkout spree → my take girl humanoids become the hands of your autonomous agents which means the win is not a cuter gait it is closing the loop from intent → plan → meters → done imagine is over the punchline is the dock assign the agent let the soft hands finish the room so back

Luella

39,709 views • 4 days ago

34,547 ELEMENTS. 34 WARNINGS. 7 DUPLICATE ELEMENTS REVIT'S OWN SCHEDULE NEVER CAUGHT. One prompt, one live Revit model, one Claude session through the MCP connection. No plugins, no exported schedules to reformat by hand. Claude walked the model tree: rooms, families, sheets, every open warning, cross-checked counts against each other, and came back with a full audit. Top flag: three fire-rated doors modeled without a fire rating parameter. The rest ranked below it, from code-adjacent to cosmetic. Still the read-only side of Revit's official MCP server, the part Autodesk shipped first, on purpose. Write access is on their own roadmap, through a separate server built specifically for it. Nothing here moved a single element. The model just got properly looked at, faster than a manual QA pass usually runs.
0:19

34,547 ELEMENTS. 34 WARNINGS. 7 DUPLICATE ELEMENTS REVIT'S OWN SCHEDULE NEVER CAUGHT. One prompt, one live Revit model, one Claude session through the MCP connection. No plugins, no exported schedules to reformat by hand. Claude walked the model tree: rooms, families, sheets, every open warning, cross-checked counts against each other, and came back with a full audit. Top flag: three fire-rated doors modeled without a fire rating parameter. The rest ranked below it, from code-adjacent to cosmetic. Still the read-only side of Revit's official MCP server, the part Autodesk shipped first, on purpose. Write access is on their own roadmap, through a separate server built specifically for it. Nothing here moved a single element. The model just got properly looked at, faster than a manual QA pass usually runs.

Solvaix

28,689 views • 1 month ago

a contractor in Shenzhen priced a ¥12,470,900 hospital contract, about $1.7m, in one afternoon and beat firms carrying forty people he explained how he did it: the bid consultancy he used to pay took three days and ¥46,000 for the same envelope. he did this one alone, off one screen, at 11.4% margin, uploaded before the 17:00 cutoff 214 pages of tender documents read, 68 binding clauses pulled out, 9,485 building parts loaded, 14 places found where a duct and a beam sit in the same cubic metre, deepest one 38mm, all of them fixed, 3,318 lines of quantities priced and the package encrypted and uploaded before the 17:00 cutoff this is Graph Engineering: the job gets cut into small nodes, one narrow task each, wired so that one node's output is the next node's input, and any node is allowed to stop the whole run. it turns a model that answers you into a machine that finishes the job: - give every node one job and one output. a node doing two things fails at both and you cannot tell which one broke - put the cheapest rejection first. his qualification node reads clause 7.4, foreign-owned firms barred, and ends the run four seconds in, before anything expensive touches the model - what moves between nodes is a file. the model travels as a model, the quantities as a table, the price as a number - build exactly one loop: the checker finds 14 collisions, the fixer drops the duct 550mm, the checker runs again, and nothing moves on until the count is zero - cap that loop, or a graph will grind on three impossible clashes until the deadline passes - keep one node whose only job is to say no, and give it authority over everything above it - log each node's output on its own, because when the price comes out wrong you need to know which node believed the wrong thing - run the expensive nodes last, always the catch is that a graph is an extremely confident machine: point it at an outdated rate book and it prices an entire hospital off it without a single node noticing, because no node is asked to doubt the input, only to process it so the nodes that earn their keep are the ones that reject, and almost nobody builds those first bookmark this, the full build with all nine nodes and what each one hands to the next is written out in the article ↓
0:16

a contractor in Shenzhen priced a ¥12,470,900 hospital contract, about $1.7m, in one afternoon and beat firms carrying forty people he explained how he did it: the bid consultancy he used to pay took three days and ¥46,000 for the same envelope. he did this one alone, off one screen, at 11.4% margin, uploaded before the 17:00 cutoff 214 pages of tender documents read, 68 binding clauses pulled out, 9,485 building parts loaded, 14 places found where a duct and a beam sit in the same cubic metre, deepest one 38mm, all of them fixed, 3,318 lines of quantities priced and the package encrypted and uploaded before the 17:00 cutoff this is Graph Engineering: the job gets cut into small nodes, one narrow task each, wired so that one node's output is the next node's input, and any node is allowed to stop the whole run. it turns a model that answers you into a machine that finishes the job: - give every node one job and one output. a node doing two things fails at both and you cannot tell which one broke - put the cheapest rejection first. his qualification node reads clause 7.4, foreign-owned firms barred, and ends the run four seconds in, before anything expensive touches the model - what moves between nodes is a file. the model travels as a model, the quantities as a table, the price as a number - build exactly one loop: the checker finds 14 collisions, the fixer drops the duct 550mm, the checker runs again, and nothing moves on until the count is zero - cap that loop, or a graph will grind on three impossible clashes until the deadline passes - keep one node whose only job is to say no, and give it authority over everything above it - log each node's output on its own, because when the price comes out wrong you need to know which node believed the wrong thing - run the expensive nodes last, always the catch is that a graph is an extremely confident machine: point it at an outdated rate book and it prices an entire hospital off it without a single node noticing, because no node is asked to doubt the input, only to process it so the nodes that earn their keep are the ones that reject, and almost nobody builds those first bookmark this, the full build with all nine nodes and what each one hands to the next is written out in the article ↓

Argona

38,189 views • 1 month ago

Unlimited-OCR is a 3B parameter model that parses entire 100-page PDFs in one shot — no page chunking, no lost context. +⁠32K context window, reads the whole document in a single pass +⁠93% on standard parsing benchmarks, +6 over baseline +⁠Under 0.11 error rate past page 40 — where every other OCR tool falls apart ⁠+Multilingual out of the box, runs locally via Transformers, Ollama, llama.cpp, Docker +⁠Textract, Google Vision, and Azure Doc Intelligence charge $1.50–$15 per 1,000 pages 1.9M downloads on HuggingFace and most people have never heard of it. This runs on your machine. For free. Forever. Model link:
0:13

Unlimited-OCR is a 3B parameter model that parses entire 100-page PDFs in one shot — no page chunking, no lost context. +⁠32K context window, reads the whole document in a single pass +⁠93% on standard parsing benchmarks, +6 over baseline +⁠Under 0.11 error rate past page 40 — where every other OCR tool falls apart ⁠+Multilingual out of the box, runs locally via Transformers, Ollama, llama.cpp, Docker +⁠Textract, Google Vision, and Azure Doc Intelligence charge $1.50–$15 per 1,000 pages 1.9M downloads on HuggingFace and most people have never heard of it. This runs on your machine. For free. Forever. Model link:

0xMarioNawfal

84,468 views • 1 month ago

$640,000 of humanoid robots died in 6 seconds because nobody ever shipped the code for running away. 9 men. Wooden handles. 40 machines that kept walking into the swing. That's the story of this clip. Not the violence. The gait. The column keeps walking because walking is all that stack does. Here's what's actually inside one of those bodies. The legs. 12 of the 43 joints live below the waist. Each knee runs a harmonic-drive or planetary actuator — a $600 to $2,000 part, sealed, non-serviceable in the field. One clean hit on a knee housing ends the unit. Not the software. The gearbox. The head. On most platforms that shell holds a depth camera and a LiDAR puck - around $250 for a RealSense, $500 to $700 for the LiDAR. Take the head off and the body doesn't die. It keeps balancing on IMU and joint encoders alone. That's why decapitated units in the clip stay upright for another 2 steps. The controller. Balance runs at 500 to 1,000 Hz. Perception runs at 30 frames a second. Those are different worlds. The balance loop is fast enough to catch a shove; the perception loop is slow, and it was trained on floors, boxes, doors, and stairs. A man sprinting in from 4 meters with a wooden handle isn't in the dataset. There's no class for it. Fall recovery exists. Every serious platform has it - G1 stands itself up, Atlas rolls and rises. Threat response exists on nothing that ships. Nobody sells it. Nobody's asked for it. Now the money. 40 units at $16,000 is $640,000 in hardware. 6 seconds of swinging takes out 60% of it. Actuators, shells, sensor stacks. The batteries - 9,000 mAh, 2 to 4 hours of walk time - are the part you don't want cracked open on a wet street. And the law is a blank page. In the US, smashing one is criminal mischief: property damage, valued at replacement cost. Same statute as a mailbox. No jurisdiction on earth has a separate line for it. The 4 known Spot attacks since 2019 all closed as vandalism. So the brief for the next generation writes itself. Not weapons, not defense. Cheaper knees, ruggedized shells, and a perception model that has finally seen a person running at it. 40 units, 43 joints each, 1,720 things to break. They didn't fail to fight back. Nobody shipped that feature.
0:14

$640,000 of humanoid robots died in 6 seconds because nobody ever shipped the code for running away. 9 men. Wooden handles. 40 machines that kept walking into the swing. That's the story of this clip. Not the violence. The gait. The column keeps walking because walking is all that stack does. Here's what's actually inside one of those bodies. The legs. 12 of the 43 joints live below the waist. Each knee runs a harmonic-drive or planetary actuator — a $600 to $2,000 part, sealed, non-serviceable in the field. One clean hit on a knee housing ends the unit. Not the software. The gearbox. The head. On most platforms that shell holds a depth camera and a LiDAR puck - around $250 for a RealSense, $500 to $700 for the LiDAR. Take the head off and the body doesn't die. It keeps balancing on IMU and joint encoders alone. That's why decapitated units in the clip stay upright for another 2 steps. The controller. Balance runs at 500 to 1,000 Hz. Perception runs at 30 frames a second. Those are different worlds. The balance loop is fast enough to catch a shove; the perception loop is slow, and it was trained on floors, boxes, doors, and stairs. A man sprinting in from 4 meters with a wooden handle isn't in the dataset. There's no class for it. Fall recovery exists. Every serious platform has it - G1 stands itself up, Atlas rolls and rises. Threat response exists on nothing that ships. Nobody sells it. Nobody's asked for it. Now the money. 40 units at $16,000 is $640,000 in hardware. 6 seconds of swinging takes out 60% of it. Actuators, shells, sensor stacks. The batteries - 9,000 mAh, 2 to 4 hours of walk time - are the part you don't want cracked open on a wet street. And the law is a blank page. In the US, smashing one is criminal mischief: property damage, valued at replacement cost. Same statute as a mailbox. No jurisdiction on earth has a separate line for it. The 4 known Spot attacks since 2019 all closed as vandalism. So the brief for the next generation writes itself. Not weapons, not defense. Cheaper knees, ruggedized shells, and a perception model that has finally seen a person running at it. 40 units, 43 joints each, 1,720 things to break. They didn't fail to fight back. Nobody shipped that feature.

HodlReaper

130,717 views • 6 days ago

A good technical LLM interview question: Your LLM chatbot takes 12s before it generates the first token, and the users are complaining. So you move the model onto a GPU with 3x the computing power. The time to first token barely improves. Why did this happen? (answer below) Latency in an LLM app is a placement problem disguised as a model problem. If you profile the 12 seconds, the model's prefill itself may only account for around 1.5 seconds of it. So halving the prefill step saves just 750ms out of 12000, which is under 7%. The rest is spread across stages that never touch the GPU. The request first travels to whatever region the app runs in, and a cross-continent round trip could cost over a second before any code executes. Then the request handler starts. On a container-based serverless platform under load, this adds several seconds of cold start, paid before auth, rate limiting, or prompt assembly even begins. Retrieval adds its own hop, and the response streams back across the same distance. Optimizing a stage that was already fast cannot alter the latency that's majorly affected by other stages. Those other stages are slow for a structural reason. An LLM app runs two workloads that want opposite machines. - The request path is short, spiky, and needs to sit close to users - Inference is long-running, GPU-bound, and billed hourly, whether requests arrive or not. So the actual decision is not which model to run, but where each of these two workloads runs. There are three options, each with its own tradeoffs: > A dedicated GPU box removes inference cold starts, but it bills around the clock and lives in one location, so distant users wait out the round trip on every request > Container-based serverless scales to zero, but the request path pays a cold start, and most of these platforms have no GPU behind them. > Edge runtimes start in under a millisecond, because a WebAssembly module carries no OS or container image to boot. They handle the request path well and cannot hold a model. So the answer is not to pick one, but to split the app across two of them. The request path runs close to users, and inference runs on a dedicated GPU it calls into. That also explains the failed upgrade. More compute made a stage that was already fast faster, and left the 10.5 seconds around it untouched. To actually learn how it's done in practice, Akamai's GitHub has a reference implementation for each half. - vllm-on-lke serves Qwen2.5-7B-Instruct behind an OpenAI-compatible endpoint on one RTX 4000 Ada GPU in Linode Kubernetes Engine, with Terraform creating the cluster, both firewalls, and the GPU operator in one apply. - akamai-functions-llm-chatbot covers the front, where a WebAssembly API checks a KV cache and only calls the GPU-backed instance on a miss. Both are available on Akamai’s new Developer Hub, alongside their tutorials and code samples. It also links to Edge Case, their Discord, where four developer advocates architect and deploy a production app live every other Wednesday. If you create a new Akamai Cloud account, you can also get $300 in credits for joining. Join here: That said, this post treats generation as a single 1.5s block, but that block has its own structure, and knowing it well tells you whether a model is slow to start or slow to stream. I wrote a first-principles walkthrough of it, covering the prefill and decode split, KV caching, and where the time actually goes inside each one. Read it below. Thanks to Akamai Cloud for partnering today!
0:18

A good technical LLM interview question: Your LLM chatbot takes 12s before it generates the first token, and the users are complaining. So you move the model onto a GPU with 3x the computing power. The time to first token barely improves. Why did this happen? (answer below) Latency in an LLM app is a placement problem disguised as a model problem. If you profile the 12 seconds, the model's prefill itself may only account for around 1.5 seconds of it. So halving the prefill step saves just 750ms out of 12000, which is under 7%. The rest is spread across stages that never touch the GPU. The request first travels to whatever region the app runs in, and a cross-continent round trip could cost over a second before any code executes. Then the request handler starts. On a container-based serverless platform under load, this adds several seconds of cold start, paid before auth, rate limiting, or prompt assembly even begins. Retrieval adds its own hop, and the response streams back across the same distance. Optimizing a stage that was already fast cannot alter the latency that's majorly affected by other stages. Those other stages are slow for a structural reason. An LLM app runs two workloads that want opposite machines. - The request path is short, spiky, and needs to sit close to users - Inference is long-running, GPU-bound, and billed hourly, whether requests arrive or not. So the actual decision is not which model to run, but where each of these two workloads runs. There are three options, each with its own tradeoffs: > A dedicated GPU box removes inference cold starts, but it bills around the clock and lives in one location, so distant users wait out the round trip on every request > Container-based serverless scales to zero, but the request path pays a cold start, and most of these platforms have no GPU behind them. > Edge runtimes start in under a millisecond, because a WebAssembly module carries no OS or container image to boot. They handle the request path well and cannot hold a model. So the answer is not to pick one, but to split the app across two of them. The request path runs close to users, and inference runs on a dedicated GPU it calls into. That also explains the failed upgrade. More compute made a stage that was already fast faster, and left the 10.5 seconds around it untouched. To actually learn how it's done in practice, Akamai's GitHub has a reference implementation for each half. - vllm-on-lke serves Qwen2.5-7B-Instruct behind an OpenAI-compatible endpoint on one RTX 4000 Ada GPU in Linode Kubernetes Engine, with Terraform creating the cluster, both firewalls, and the GPU operator in one apply. - akamai-functions-llm-chatbot covers the front, where a WebAssembly API checks a KV cache and only calls the GPU-backed instance on a miss. Both are available on Akamai’s new Developer Hub, alongside their tutorials and code samples. It also links to Edge Case, their Discord, where four developer advocates architect and deploy a production app live every other Wednesday. If you create a new Akamai Cloud account, you can also get $300 in credits for joining. Join here: That said, this post treats generation as a single 1.5s block, but that block has its own structure, and knowing it well tells you whether a model is slow to start or slow to stream. I wrote a first-principles walkthrough of it, covering the prefill and decode split, KV caching, and where the time actually goes inside each one. Read it below. Thanks to Akamai Cloud for partnering today!

Avi Chawla

21,786 views • 25 days ago

THE MEME SIDE OF MY GROK BOT MADE $2,014 TODAY. THE PREDICTION SIDE MADE $297. SAME BOT, SAME AFTERNOON the account was $89 five days ago and closed at $4,339.39 this morning. the split is the whole story most people point a bot at one market and let it guess. a bot with one input is a chatbot with a wallet. mine runs five parts, in this order: → TRENCH. Robinhood Chain memecoins, the main engine now. $1,224.63 went in this morning and came back as $3,239.11. four positions opened, three closed while the chart still looked healthy → MAP. a free open source world map, zero API keys, running on my own machine. conflicts, shipping lanes, cables, weather, all live. attention lands there hours before it lands on a chart → RESEARCH. before a dollar moves it checks the story against primary sources. nineteen launches got skipped today on that check or the liquidity filter → ODDS. prediction markets are the slow side now. seven books, all closed today, $803.12 came back as $1,100.27 → EXIT. one rule closes everything: six hour volume under 20% of the 24 hour average. no discretion, no averaging down the exit rule is the part nobody copies, and the only reason the meme side is not a donation every launch dies eventually. the only question is whether you are still holding when the volume goes TRENCH → MAP → RESEARCH → ODDS → EXIT. everyone else is still typing one off prompts and wondering why nothing compounds Bookmark this before your next entry the map is open source and runs with zero API keys. say the word in the replies and I will send the link
0:19

THE MEME SIDE OF MY GROK BOT MADE $2,014 TODAY. THE PREDICTION SIDE MADE $297. SAME BOT, SAME AFTERNOON the account was $89 five days ago and closed at $4,339.39 this morning. the split is the whole story most people point a bot at one market and let it guess. a bot with one input is a chatbot with a wallet. mine runs five parts, in this order: → TRENCH. Robinhood Chain memecoins, the main engine now. $1,224.63 went in this morning and came back as $3,239.11. four positions opened, three closed while the chart still looked healthy → MAP. a free open source world map, zero API keys, running on my own machine. conflicts, shipping lanes, cables, weather, all live. attention lands there hours before it lands on a chart → RESEARCH. before a dollar moves it checks the story against primary sources. nineteen launches got skipped today on that check or the liquidity filter → ODDS. prediction markets are the slow side now. seven books, all closed today, $803.12 came back as $1,100.27 → EXIT. one rule closes everything: six hour volume under 20% of the 24 hour average. no discretion, no averaging down the exit rule is the part nobody copies, and the only reason the meme side is not a donation every launch dies eventually. the only question is whether you are still holding when the volume goes TRENCH → MAP → RESEARCH → ODDS → EXIT. everyone else is still typing one off prompts and wondering why nothing compounds Bookmark this before your next entry the map is open source and runs with zero API keys. say the word in the replies and I will send the link

savip.

51,859 views • 13 days ago

i gave bonsai, 3.9 gb model one spec to build a working rotary engine autonomously, using hermes agent. all on rtx 3060ti 8gb. watch it go from a black canvas to a running wankel. the epitrochoid housing, a triangular rotor turning at exactly a third of shaft speed, three chambers firing intake, compression, power, exhaust, and ten geometry tests it had to pass to count as done. a provably correct simulation, not a drawing. then i left it alone and it worked. wrote the code, ran its own tests, hit failures, read them, fixed its own lines, until the engine spun and all ten went green. a real agent loop, and i watched every step. it broke three times getting here and clawed back every one. that honest part is in the reply. and this is a sneak. i've been building a show around runs exactly like this, and it's almost ready. more very soon. just getting started.
0:17

i gave bonsai, 3.9 gb model one spec to build a working rotary engine autonomously, using hermes agent. all on rtx 3060ti 8gb. watch it go from a black canvas to a running wankel. the epitrochoid housing, a triangular rotor turning at exactly a third of shaft speed, three chambers firing intake, compression, power, exhaust, and ten geometry tests it had to pass to count as done. a provably correct simulation, not a drawing. then i left it alone and it worked. wrote the code, ran its own tests, hit failures, read them, fixed its own lines, until the engine spun and all ten went green. a real agent loop, and i watched every step. it broke three times getting here and clawed back every one. that honest part is in the reply. and this is a sneak. i've been building a show around runs exactly like this, and it's almost ready. more very soon. just getting started.

Sudo su

34,883 views • 1 month ago

Stop renting a chatbot and calling it your company’s brain. Utopia is the first open-source enterprise world model I’ve seen that actually treats knowledge like a system, not a vibe. It's a Self-hosted, open source and built around the one question every other knowledge base ignores: when. Every fact in it carries a start date and an end date. Nothing gets overwritten. When something changes, the correction closes the old version and opens a new one. The history stays intact. So the graph isn't a snapshot anymore. It's a recording you can rewind. You drag a timeline on the screen and watch the whole thing redraw itself to show what was true on that date. Who owned what. What the policy said before it changed. All of it, replayable. And every fact points back to the exact sentence it came from. Nothing gets in without a receipt. Here's the part that got me. The whole thing is one binary and one Postgres database. That's it. No Elasticsearch, no separate vector service and no message queue. Most RAG stacks are six services duct-taped together. This is two. It even runs fully offline on an air-gapped network with any model you want.
0:16

Stop renting a chatbot and calling it your company’s brain. Utopia is the first open-source enterprise world model I’ve seen that actually treats knowledge like a system, not a vibe. It's a Self-hosted, open source and built around the one question every other knowledge base ignores: when. Every fact in it carries a start date and an end date. Nothing gets overwritten. When something changes, the correction closes the old version and opens a new one. The history stays intact. So the graph isn't a snapshot anymore. It's a recording you can rewind. You drag a timeline on the screen and watch the whole thing redraw itself to show what was true on that date. Who owned what. What the policy said before it changed. All of it, replayable. And every fact points back to the exact sentence it came from. Nothing gets in without a receipt. Here's the part that got me. The whole thing is one binary and one Postgres database. That's it. No Elasticsearch, no separate vector service and no message queue. Most RAG stacks are six services duct-taped together. This is two. It even runs fully offline on an air-gapped network with any model you want.

Hasan Toor

173,779 views • 11 days ago

New open source library: kugiri (区切り). Already running in production on several of my sites, now public. It splits text into lines, words and characters, the thing every text reveal needs. The difference is where it cuts. Most splitters guess where a line breaks, and the guess falls apart the moment the text does something real: text-wrap: balance, hyphenation, a drop cap, a float, a link that wraps, RTL, Japanese. kugiri does not guess. It asks the browser where the lines already are and cuts exactly there, so the split text looks pixel for pixel like the text before it. No animation, no scroll watching, no opinions. Every unit gets an index and the library gets out of the way. Animate with CSS, the Web Animations API, GSAP or Motion. One file, no dependencies, about 7 kB gzipped. The demo is also the test suite: fifty situations, each checked against what the browser painted, in Chromium, WebKit and Firefox. Demo: npm: Source:
0:17

New open source library: kugiri (区切り). Already running in production on several of my sites, now public. It splits text into lines, words and characters, the thing every text reveal needs. The difference is where it cuts. Most splitters guess where a line breaks, and the guess falls apart the moment the text does something real: text-wrap: balance, hyphenation, a drop cap, a float, a link that wraps, RTL, Japanese. kugiri does not guess. It asks the browser where the lines already are and cuts exactly there, so the split text looks pixel for pixel like the text before it. No animation, no scroll watching, no opinions. Every unit gets an index and the library gets out of the way. Animate with CSS, the Web Animations API, GSAP or Motion. One file, no dependencies, about 7 kB gzipped. The demo is also the test suite: fifty situations, each checked against what the browser painted, in Chromium, WebKit and Firefox. Demo: npm: Source:

Edoardo Lunardi

42,898 views • 8 days ago

Stanford researchers did it again. They just built the agent-native version of Git. When an agent works on a longer task, the run builds up a lot of state. This includes files edited/created, a dev server, a database, installed packages, KV cache, etc. Say the agent is at step 10 and makes a mistake, maybe it misreads a traceback and rewrites a file that was actually fine. The tests start failing, and the run goes off track, although everything through step eight was correct. By default, the agent just tries to fix it, which creates more edits and tool calls. This burns more tokens and grows the context. The other options are a person stepping in to redirect it or restarting the whole run from step one. That's wasteful, because it pays for every model/tool call again and re-prefills the context. Moreover, since an agent's run is non-deterministic, it doesn't reproduce the same early steps anyway. The reason it's hard to just jump back exactly to a previous correct step and resume from there is that the trajectory is only a message log. It records what the agent said and which tools it called, but not the live state underneath. That state includes things like memory, open file handles, child processes, installed packages, /tmp, and KV cache. None of that is in the log. Git can version the files, but it doesn't snapshot the running process or the KV cache. Checking out step eight moves the files back, but the process is still sitting in step-ten memory with a cold cache. Shepherd is a runtime layer by Stanford that records the run as a trace of typed events rather than a flat log. Each agent-environment interaction becomes a commit, similar to Git, but it tracks the live run. Its commit includes the agent process and the filesystem together, copy-on-write, so a branch carries the actual state and not just the files. Going back to a previous step is then a single call that forks from that commit and continues from the exact state. The copy-on-write fork is roughly five times faster than docker commit, and because the prompt prefix through step eight is unchanged, the KV cache is reused over 95% on replay, so early steps aren't reprocessed again. Once the run can be forked, a meta-agent can sit on top and operate it. It watches the trace and reverts as soon as it looks wrong, before the bad write is committed. In practice, it's just Python calling fork, replay, and revert on the trace, rather than a separate control plane wired into the harness. Not everything is reversible though. Files and sandbox changes undo themselves, but a database write has no automatic undo, so it needs a matching undo step set up in advance. Something external, like a sent email or a real charge, can't be undone, so the supervisor's job there is to catch it before it fires. They tested this on a few public benchmarks. On CooperBench, where two agents work on the same codebase, adding a live supervisor took the pair-coding pass rate from 28.8% to 54.7%. It's still early and labeled alpha. The benefit mostly shows up when a run gets branched a lot over a heavy sandbox state, which is exactly where restarting wastes the most tokens and time. If Git was made to make file changes reversible, Shepherd is trying to do the same thing for a live agent run. Shepherd Repo: (don't forget to star it ⭐ ) That said, Shepherd reverts a bad step inside a run. The harness around it, the prompts, tools, and checks the supervisor relies on, still drifts across runs as models and dependencies change. Akshay wrote about making that harness repair itself, where a failing trace gets diagnosed, the fix is verified against the exact input that failed, and the failure is locked as a regression test so it can't recur. Read it below.
0:15

Stanford researchers did it again. They just built the agent-native version of Git. When an agent works on a longer task, the run builds up a lot of state. This includes files edited/created, a dev server, a database, installed packages, KV cache, etc. Say the agent is at step 10 and makes a mistake, maybe it misreads a traceback and rewrites a file that was actually fine. The tests start failing, and the run goes off track, although everything through step eight was correct. By default, the agent just tries to fix it, which creates more edits and tool calls. This burns more tokens and grows the context. The other options are a person stepping in to redirect it or restarting the whole run from step one. That's wasteful, because it pays for every model/tool call again and re-prefills the context. Moreover, since an agent's run is non-deterministic, it doesn't reproduce the same early steps anyway. The reason it's hard to just jump back exactly to a previous correct step and resume from there is that the trajectory is only a message log. It records what the agent said and which tools it called, but not the live state underneath. That state includes things like memory, open file handles, child processes, installed packages, /tmp, and KV cache. None of that is in the log. Git can version the files, but it doesn't snapshot the running process or the KV cache. Checking out step eight moves the files back, but the process is still sitting in step-ten memory with a cold cache. Shepherd is a runtime layer by Stanford that records the run as a trace of typed events rather than a flat log. Each agent-environment interaction becomes a commit, similar to Git, but it tracks the live run. Its commit includes the agent process and the filesystem together, copy-on-write, so a branch carries the actual state and not just the files. Going back to a previous step is then a single call that forks from that commit and continues from the exact state. The copy-on-write fork is roughly five times faster than docker commit, and because the prompt prefix through step eight is unchanged, the KV cache is reused over 95% on replay, so early steps aren't reprocessed again. Once the run can be forked, a meta-agent can sit on top and operate it. It watches the trace and reverts as soon as it looks wrong, before the bad write is committed. In practice, it's just Python calling fork, replay, and revert on the trace, rather than a separate control plane wired into the harness. Not everything is reversible though. Files and sandbox changes undo themselves, but a database write has no automatic undo, so it needs a matching undo step set up in advance. Something external, like a sent email or a real charge, can't be undone, so the supervisor's job there is to catch it before it fires. They tested this on a few public benchmarks. On CooperBench, where two agents work on the same codebase, adding a live supervisor took the pair-coding pass rate from 28.8% to 54.7%. It's still early and labeled alpha. The benefit mostly shows up when a run gets branched a lot over a heavy sandbox state, which is exactly where restarting wastes the most tokens and time. If Git was made to make file changes reversible, Shepherd is trying to do the same thing for a live agent run. Shepherd Repo: (don't forget to star it ⭐ ) That said, Shepherd reverts a bad step inside a run. The harness around it, the prompts, tools, and checks the supervisor relies on, still drifts across runs as models and dependencies change. Akshay wrote about making that harness repair itself, where a failing trace gets diagnosed, the fix is verified against the exact input that failed, and the failure is locked as a regression test so it can't recur. Read it below.

Avi Chawla

441,703 views • 2 months ago