Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

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.show more

Akshay 🚀

289,784 subscribers

245,379 просмотров • 3 месяцев назад •via X (Twitter)

Наука и технологии

Anya Rossi• Live Now

Private livecam show

Комментарии: 0

Нет доступных комментариев

Здесь появятся комментарии из оригинального поста

Похожие видео

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.

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 просмотров • 1 месяц назад

If you are trying to understand where AI agents are going, learn harness engineering. A capable model is only one part of an agent system. Once the model begins reading files, calling tools, modifying state and working across many steps, the quality of the system depends increasingly on the software around it. Consider a coding agent working through a large repository. The model can decide that it needs to inspect a file, search for a symbol, make an edit or run a test, but those decisions do not execute themselves. The surrounding runtime has to decide which resources are available, whether the requested action is permitted, how the operation should be performed, what result should be retained, and what information should be presented to the model on the next step. This becomes harder as the run gets longer. As history accumulates, replaying everything can become costly and less effective. The harness has to decide what should remain in context, what should be summarized or retrieved later, and what belongs in persistent state outside the context window. Execution has similar problems. A long-running agent may need to survive an interruption, avoid repeating completed work, enforce permissions around consequential actions, and preserve enough history to reconstruct what happened when the final result is wrong. These are harness problems. The harness is the layer that manages context, tools, execution, state, checkpoints, limits and traces around the model. Harness engineering is the work of designing and improving that layer. Engineers inspect execution traces, evaluate agents on representative tasks, look for recurring failure modes, and then change things such as context selection, tool interfaces, state handling or execution controls. That last part matters because agent failures are often not fixed by changing the model. Sometimes the useful change is in what the model sees, how a tool is exposed, what state is preserved, or what the runtime does after a failed step. As agents take on longer tasks, the demands on this surrounding software grow. Model capability remains essential, but harness engineering is what turns that capability into an execution process that can be controlled, inspected, tested and improved.

If you are trying to understand where AI agents are going, learn harness engineering. A capable model is only one part of an agent system. Once the model begins reading files, calling tools, modifying state and working across many steps, the quality of the system depends increasingly on the software around it. Consider a coding agent working through a large repository. The model can decide that it needs to inspect a file, search for a symbol, make an edit or run a test, but those decisions do not execute themselves. The surrounding runtime has to decide which resources are available, whether the requested action is permitted, how the operation should be performed, what result should be retained, and what information should be presented to the model on the next step. This becomes harder as the run gets longer. As history accumulates, replaying everything can become costly and less effective. The harness has to decide what should remain in context, what should be summarized or retrieved later, and what belongs in persistent state outside the context window. Execution has similar problems. A long-running agent may need to survive an interruption, avoid repeating completed work, enforce permissions around consequential actions, and preserve enough history to reconstruct what happened when the final result is wrong. These are harness problems. The harness is the layer that manages context, tools, execution, state, checkpoints, limits and traces around the model. Harness engineering is the work of designing and improving that layer. Engineers inspect execution traces, evaluate agents on representative tasks, look for recurring failure modes, and then change things such as context selection, tool interfaces, state handling or execution controls. That last part matters because agent failures are often not fixed by changing the model. Sometimes the useful change is in what the model sees, how a tool is exposed, what state is preserved, or what the runtime does after a failed step. As agents take on longer tasks, the demands on this surrounding software grow. Model capability remains essential, but harness engineering is what turns that capability into an execution process that can be controlled, inspected, tested and improved.

Tech with Mak

48,318 просмотров • 8 дней назад

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.

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

31,198 просмотров • 1 месяц назад

Big win for open-source LLMs! DeepSeek V4 Pro holds the top open-weights score on SWE-bench Verified, in the GPT-5.5 range. GLM 5.2 leads the open-weight intelligence index and sits near the closed frontier on long-horizon coding. But this leaderboard number is a weak proxy for real performance. It comes from one task set, run through one harness, served at one precision. The same weights can even score differently across providers, since many hosts quantize activations to fp8 and drift the model off its reference weights. Real performance is determined based on whether a model can read a repo, make coordinated edits across files, run the tests, and recover when one breaks. By that measure, the top open models hold up, but only inside the right harness. The teams that actually put DeepSeek V4 into production pipelines as a frontier substitute got there through the harness they built around the model, not by picking a stronger model. If you want to see this in practice, Cline (64k+ stars) has actually built that harness around open models, tuned so they run at production quality. And it's tuned so that these LLMs can run at production quality, with plan and act modes, checkpoints, and terminal feedback. ClinePass is the new access layer on top of it. It runs a curated set of those models inside Cline, narrowed to the ones tested for coding-agent use, with 2 to 5x the standard rate limits and no separate provider accounts, keys, or billing to track. The video below shows the setup, and I worked with the team to put this together. It runs alongside custom keys and local models as well, not in place of them.

Big win for open-source LLMs! DeepSeek V4 Pro holds the top open-weights score on SWE-bench Verified, in the GPT-5.5 range. GLM 5.2 leads the open-weight intelligence index and sits near the closed frontier on long-horizon coding. But this leaderboard number is a weak proxy for real performance. It comes from one task set, run through one harness, served at one precision. The same weights can even score differently across providers, since many hosts quantize activations to fp8 and drift the model off its reference weights. Real performance is determined based on whether a model can read a repo, make coordinated edits across files, run the tests, and recover when one breaks. By that measure, the top open models hold up, but only inside the right harness. The teams that actually put DeepSeek V4 into production pipelines as a frontier substitute got there through the harness they built around the model, not by picking a stronger model. If you want to see this in practice, Cline (64k+ stars) has actually built that harness around open models, tuned so they run at production quality. And it's tuned so that these LLMs can run at production quality, with plan and act modes, checkpoints, and terminal feedback. ClinePass is the new access layer on top of it. It runs a curated set of those models inside Cline, narrowed to the ones tested for coding-agent use, with 2 to 5x the standard rate limits and no separate provider accounts, keys, or billing to track. The video below shows the setup, and I worked with the team to put this together. It runs alongside custom keys and local models as well, not in place of them.

Avi Chawla

44,124 просмотров • 3 месяцев назад

this is straight f*cking gold the full harness guide for Kimi K3, the #1 open source frontend model the premise: the model underneath keeps changing. the harness is the part that stays yours what one night of it looks like: 02:00 - the trigger fires, every node that needs work gets picked > 212 agents fan out, one per node > a bad return gets rejected, retried once with the reason attached > one agent tries to write outside its folder. blocked. nobody woken up > +41 nodes and +96 edges land in the graph > a drafted email hits pre_send and waits for you 02:41 - the loop stops on its own, inside a 45 minute budget 07:30 - you read one file and make two decisions the whole machine is one folder, one config file, five short scripts two rules hold it together: > the model sits behind one line, so a better model is a config change > the verifier lives outside the agent, so nothing grades its own work now the money part companies burn whole quarters building internal agent platforms a harness setup for a small team goes for four figures, one time then a monthly retainer to run the swap test on every new model release models come and go. the person who owns the harness keeps getting paid

this is straight f*cking gold the full harness guide for Kimi K3, the #1 open source frontend model the premise: the model underneath keeps changing. the harness is the part that stays yours what one night of it looks like: 02:00 - the trigger fires, every node that needs work gets picked > 212 agents fan out, one per node > a bad return gets rejected, retried once with the reason attached > one agent tries to write outside its folder. blocked. nobody woken up > +41 nodes and +96 edges land in the graph > a drafted email hits pre_send and waits for you 02:41 - the loop stops on its own, inside a 45 minute budget 07:30 - you read one file and make two decisions the whole machine is one folder, one config file, five short scripts two rules hold it together: > the model sits behind one line, so a better model is a config change > the verifier lives outside the agent, so nothing grades its own work now the money part companies burn whole quarters building internal agent platforms a harness setup for a small team goes for four figures, one time then a monthly retainer to run the swap test on every new model release models come and go. the person who owns the harness keeps getting paid

Mr. Buzzoni

12,633 просмотров • 17 дней назад

50% cheaper Claude inference with just one line of code change! - Remove → model="claude-opus-4-8" - Add → model="ship-like/claude-opus-4-8" I verified the cost saving in my own terminal by invoking the same Anthropic model with the same prompt. The underlying engineering by Ship is actually interesting, and the patterns can be used in any production LLM stack. Essentially, a trained model is a frozen artifact. Every request performs the same forward-pass, whether it extracts a date or refactors a module, because the compute decision was made at training time, before the request existed. Ship makes that decision at inference time instead. After seeing a request, it searches over executions, involving single models, cascades, ensembles, or harnesses with tools, and serves the cheapest one that will match the reference model's quality. This is not a basic router, because picking a cheaper model per query doesn't ensure the cheaper model preserves the original's behavior, like output shape, tool-call patterns, and refusals. Ship measures this equivalence directly. Outputs stay distributionally indistinguishable from the reference model, not token-identical, since two calls to the same model already differ, but they are indistinguishable in capability and behavior. Of course, some requests execute cheaply and some cost Ship more than the customer pays, but the price per request is still a flat 50% off either way, so the execution-cost variance moves off the application's bill entirely. The video below depicts the cost savings and output in my real invocation, and I partnered with the team to put this together.

50% cheaper Claude inference with just one line of code change! - Remove → model="claude-opus-4-8" - Add → model="ship-like/claude-opus-4-8" I verified the cost saving in my own terminal by invoking the same Anthropic model with the same prompt. The underlying engineering by Ship is actually interesting, and the patterns can be used in any production LLM stack. Essentially, a trained model is a frozen artifact. Every request performs the same forward-pass, whether it extracts a date or refactors a module, because the compute decision was made at training time, before the request existed. Ship makes that decision at inference time instead. After seeing a request, it searches over executions, involving single models, cascades, ensembles, or harnesses with tools, and serves the cheapest one that will match the reference model's quality. This is not a basic router, because picking a cheaper model per query doesn't ensure the cheaper model preserves the original's behavior, like output shape, tool-call patterns, and refusals. Ship measures this equivalence directly. Outputs stay distributionally indistinguishable from the reference model, not token-identical, since two calls to the same model already differ, but they are indistinguishable in capability and behavior. Of course, some requests execute cheaply and some cost Ship more than the customer pays, but the price per request is still a flat 50% off either way, so the execution-cost variance moves off the application's bill entirely. The video below depicts the cost savings and output in my real invocation, and I partnered with the team to put this together.

Akshay 🚀

64,482 просмотров • 2 месяцев назад

AGENT ARCHITECTURE ROUTES WORK. IT DOES NOT REMEMBER WORK. THAT GAP IS WHY YOUR LOOP KEEPS FIXING THE SAME BUG TWICE. these are two different engineering problems. every agent that silently drifts is missing one of them. architecture answers what runs. harness → loop → graph. it defines the tools, the retries, the branching routes, the approval gates. context ops answer what the run knows. write → read → compress → isolate. it defines what gets saved between attempts, pulled in on read, summarized on overflow, and split across sub-agents. for two months i believed a solid harness plus a verifier loop was enough. my coding agent kept re-discovering the same test failure across retries. the loop was working. it just had nowhere to write what it had already learned. here is the decision rule: if your agent forgets across restarts, add write and read. if it stalls on long tasks, add compress. if two sub-agents step on each other, add isolate. architecture without context ops is a well-routed system with amnesia.

AGENT ARCHITECTURE ROUTES WORK. IT DOES NOT REMEMBER WORK. THAT GAP IS WHY YOUR LOOP KEEPS FIXING THE SAME BUG TWICE. these are two different engineering problems. every agent that silently drifts is missing one of them. architecture answers what runs. harness → loop → graph. it defines the tools, the retries, the branching routes, the approval gates. context ops answer what the run knows. write → read → compress → isolate. it defines what gets saved between attempts, pulled in on read, summarized on overflow, and split across sub-agents. for two months i believed a solid harness plus a verifier loop was enough. my coding agent kept re-discovering the same test failure across retries. the loop was working. it just had nowhere to write what it had already learned. here is the decision rule: if your agent forgets across restarts, add write and read. if it stalls on long tasks, add compress. if two sub-agents step on each other, add isolate. architecture without context ops is a well-routed system with amnesia.

kocer

12,740 просмотров • 1 месяц назад

I MADE MY AI AGENT 10X FASTER WITHOUT CHANGING THE MODEL not a smarter model, not a bigger context window, not another clever prompt the same kind of AI that designs vaccines for viruses we have not even met yet was spending two minutes opening the wrong files just to hand me a brief from three months ago the problem was never capability, it was the scaffolding that piled up around my agent by accident, folder by folder an agent does not think in your categories, it searches from scratch every single time, and your tidy human folders are a maze to it the fix was almost stupidly small, one index file at the root of each big folder and a few numbers in front of the folder names slowest task dropped from 2 minutes to 26 seconds, fastest ones hit 10, zero model changes capability is cheap when the scaffolding around it is broken the article breaks down the whole system in 15 minutes ↓

I MADE MY AI AGENT 10X FASTER WITHOUT CHANGING THE MODEL not a smarter model, not a bigger context window, not another clever prompt the same kind of AI that designs vaccines for viruses we have not even met yet was spending two minutes opening the wrong files just to hand me a brief from three months ago the problem was never capability, it was the scaffolding that piled up around my agent by accident, folder by folder an agent does not think in your categories, it searches from scratch every single time, and your tidy human folders are a maze to it the fix was almost stupidly small, one index file at the root of each big folder and a few numbers in front of the folder names slowest task dropped from 2 minutes to 26 seconds, fastest ones hit 10, zero model changes capability is cheap when the scaffolding around it is broken the article breaks down the whole system in 15 minutes ↓

shmidt

36,479 просмотров • 3 месяцев назад

You can now post-train a model inside your existing production harness with our platform, AC2. A production harness is a whole engineered system around the LLM, with its own context management, tools, sandboxing, and control flows. Porting that into a new training runtime can be expensive and could introduce train-test mismatch, where the policy is optimized against a simulated harness and then struggles in production. All you need to do is swap out the harness’ LLM response endpoint to one provided by AC2, and expose a lightweight protocol for AC2 to initiate and grade rollouts; the trainer handles the rest.

You can now post-train a model inside your existing production harness with our platform, AC2. A production harness is a whole engineered system around the LLM, with its own context management, tools, sandboxing, and control flows. Porting that into a new training runtime can be expensive and could introduce train-test mismatch, where the policy is optimized against a simulated harness and then struggles in production. All you need to do is swap out the harness’ LLM response endpoint to one provided by AC2, and expose a lightweight protocol for AC2 to initiate and grade rollouts; the trainer handles the rest.

Applied Compute

182,296 просмотров • 1 месяц назад

HydraFusion Explained. Part I: How does the Copilot engine know what to optimize for? Your prompt is evaluated across 4 dimensions: ➡ Does it require deep reasoning? (aka. reasoning depth) ➡ Is it a sophisticated problem? (aka. code generation complexity) ➡ Is it untangling a complicated mess? (aka. debugging difficulty) ➡ Is it dominated by tool-use? (aka. tool orchestration needs) Based on this evaluation, a HyDRA score is assigned to determine the capability profile your task needs the most and to establish a quality bar. Part II: How does it choose a model? Note: It doesn' t pick one model to handle the entire job e2e, (that's Auto mode). Instead, it selects 1 of 3 execution workflows and assigns the best model at different stages based on the HyDRA score: 1️⃣ Single ⚙️ How it works: A single model completes the task from start to finish. ⚖️ Rationale: The task comfortably meets the quality bar with one model. Multi-model orchestration would add latency and cost with no meaningful quality gain. 2️⃣ Cascade ⚙️ How it works: A lightweight, cost-efficient model generates the solution. This draft is evaluated against a quality gate and if it falls short of the quality bar, the entire task escalates to a stronger, frontier model. ⚖️ Rationale: Only bring in the big guns when there is concrete evidence that a lightweight model won't meet the quality threshold. 3️⃣ Critique ⚙️ How it works: A lightweight model drafts the initial code and tool interactions. An independent, read-only frontier model reviews that draft and provides feedback. The original lightweight model then performs any targeted revision(s) before the final response is sent to the user. ⚖️ Rationale: Writing code (output tokens) is expensive while reviewing code (input tokens) is cheap. Instead of incurring the cost of a powerhouse writing hundreds of lines from scratch, a cost-efficient model writes the first draft, and the frontier model just reviews it and points out fixes. HydraFusion is available in experimental preview on the GitHub Copilot CLI: /experimental on, /model and select Hydrafusion (Research Preview)

Julia Muiruri

12,687 просмотров • 26 дней назад

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 ↓

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

32,744 просмотров • 1 месяц назад

We released physics-intern: a simple harness for science problems! It gets models like Gemini 3.1 Pro to go from 17.7 -> 31.4, thus beating GPT 5.5 Pro. The physics-intern harness can wrap any model and via dedicated subagent boost the performance of the vanilla reasoning models. While I think more and more of these harness capability gains will be absorbed into the models (like prompting tricks disappeared over time) there is a lot to be gained right now by building good scaffolds for those models and integrating tools well. Interestingly, the exception we found that GPT 5.5 Pro actually didn't benefit from the physics-intern harness! Read more about it here: PS: I think the Harness[Model] notation is kind of nice.

We released physics-intern: a simple harness for science problems! It gets models like Gemini 3.1 Pro to go from 17.7 -> 31.4, thus beating GPT 5.5 Pro. The physics-intern harness can wrap any model and via dedicated subagent boost the performance of the vanilla reasoning models. While I think more and more of these harness capability gains will be absorbed into the models (like prompting tricks disappeared over time) there is a lot to be gained right now by building good scaffolds for those models and integrating tools well. Interestingly, the exception we found that GPT 5.5 Pro actually didn't benefit from the physics-intern harness! Read more about it here: PS: I think the Harness[Model] notation is kind of nice.

Leandro von Werra

97,519 просмотров • 4 месяцев назад

whoever leaked this has bigger balls than sense Google Research and MIT ran the same agent jobs 260 different ways for Nature last month: they held the prompts, the tools and the compute budget identical and moved nothing but the wiring between the agents, and the same work swung from 70% worse than a single agent to 80.8% better, averaging out at 0.0% i ran my own single agent against the task list first and it cleared 6 of 10 alone, already past the line where a crew starts subtracting this is Graph Engineering, the layer that decides whether a crew is worth 80% more or 70% less, and it installs into the agent you already pay for: - score your solo agent on the real task first: above roughly 45% success that study predicts zero to negative returns from any crew you put around it - under that line, put one supervisor over the fan out: crews with no correction step amplified their own errors to 17.2x the single agent rate, supervised aggregation held it to 4.4x - give every worker one output and let none of them read a peer's draft, so a wrong step reaches the supervisor instead of four other agents - run the comparison again after every model upgrade, because a better model raises your baseline and a higher baseline is what makes a crew stop paying - keep the single agent alive as the control, the only number that says the wiring is earning its calls turns out the shape does not travel: the biggest win came off a finance task under one supervisor and the worst collapse off a planning task with independent agents my position, and it is the arguable one: a crew is a bet on your own diagram, and the model you pick moves that bet less than one arrow does bookmark this, the three moves that draw those arrows before you pay for one extra call are in the post below ↓

whoever leaked this has bigger balls than sense Google Research and MIT ran the same agent jobs 260 different ways for Nature last month: they held the prompts, the tools and the compute budget identical and moved nothing but the wiring between the agents, and the same work swung from 70% worse than a single agent to 80.8% better, averaging out at 0.0% i ran my own single agent against the task list first and it cleared 6 of 10 alone, already past the line where a crew starts subtracting this is Graph Engineering, the layer that decides whether a crew is worth 80% more or 70% less, and it installs into the agent you already pay for: - score your solo agent on the real task first: above roughly 45% success that study predicts zero to negative returns from any crew you put around it - under that line, put one supervisor over the fan out: crews with no correction step amplified their own errors to 17.2x the single agent rate, supervised aggregation held it to 4.4x - give every worker one output and let none of them read a peer's draft, so a wrong step reaches the supervisor instead of four other agents - run the comparison again after every model upgrade, because a better model raises your baseline and a higher baseline is what makes a crew stop paying - keep the single agent alive as the control, the only number that says the wiring is earning its calls turns out the shape does not travel: the biggest win came off a finance task under one supervisor and the worst collapse off a planning task with independent agents my position, and it is the arguable one: a crew is a bet on your own diagram, and the model you pick moves that bet less than one arrow does bookmark this, the three moves that draw those arrows before you pay for one extra call are in the post below ↓

Argona

892,088 просмотров • 1 месяц назад

we just released a new blog "Training a coding agent using the OpenCode harness in remote HF sandboxes with TRL and OpenEnv" you can take a real coding agent (OpenCode), let it run its own tool loop against real coding problems, and train it with RL on the exact tokens it produced and every rollout runs in its own remote HF sandbox, so rollouts scale out beyond one machine the loop: - OpenCode owns its tool loop inside an OpenEnv sandbox - an in-sandbox proxy records the real token ids + logprobs, per turn - a hidden-test verifier scores the result, and that is the reward - TRL trains with AsyncGRPO, weights sync back to vLLM over NCCL blog + runnable example:

we just released a new blog "Training a coding agent using the OpenCode harness in remote HF sandboxes with TRL and OpenEnv" you can take a real coding agent (OpenCode), let it run its own tool loop against real coding problems, and train it with RL on the exact tokens it produced and every rollout runs in its own remote HF sandbox, so rollouts scale out beyond one machine the loop: - OpenCode owns its tool loop inside an OpenEnv sandbox - an in-sandbox proxy records the real token ids + logprobs, per turn - a hidden-test verifier scores the result, and that is the reward - TRL trains with AsyncGRPO, weights sync back to vLLM over NCCL blog + runnable example:

Sergio Paniego

37,936 просмотров • 2 месяцев назад

let me save you 3 hours of head scratching. if you're running local models like Qwen3.5-35B-A3B through Claude Code via llama.cpp's Anthropic endpoint, the chain will break every 3 to 5 minutes. tool call fails. flow stops. you reprompt. it recovers. 2 minutes later it stops again. the model is fine. the harness chokes on local inference latency. switch to OpenCode. same localhost endpoint. same model. same GPU. the chain doesn't break. the tradeoff: OpenCode sometimes loops. the model forgets what it already read and repeats the same tool call. but a loop you can interrupt. a broken chain kills your momentum and you start over. watch both side by side. proprietary agent vs open source agent. same 3B model. different failure modes. pick your poison.

let me save you 3 hours of head scratching. if you're running local models like Qwen3.5-35B-A3B through Claude Code via llama.cpp's Anthropic endpoint, the chain will break every 3 to 5 minutes. tool call fails. flow stops. you reprompt. it recovers. 2 minutes later it stops again. the model is fine. the harness chokes on local inference latency. switch to OpenCode. same localhost endpoint. same model. same GPU. the chain doesn't break. the tradeoff: OpenCode sometimes loops. the model forgets what it already read and repeats the same tool call. but a loop you can interrupt. a broken chain kills your momentum and you start over. watch both side by side. proprietary agent vs open source agent. same 3B model. different failure modes. pick your poison.

Sudo su

72,554 просмотров • 7 месяцев назад

𝗧𝗵𝗲 𝟰 𝗟𝗮𝘆𝗲𝗿𝘀 𝗼𝗳 𝗮𝗻 𝗔𝗴𝗲𝗻𝘁 𝗦𝘆𝘀𝘁𝗲𝗺 An agent can burn tokens, declare success, and still fail the tests. That’s often an architecture problem, not a prompting problem. 𝟭. 𝗟𝗼𝗼𝗽 → Decides whether to continue Acts → verifies → retries until there’s evidence of success. 𝟮. 𝗚𝗿𝗮𝗽𝗵 → Decides what runs next Handles branches, retries, handoffs, fallbacks, and shared state. 𝟯. 𝗛𝗮𝗿𝗻𝗲𝘀𝘀 → Gives the model its capabilities Tools, APIs, files, memory, permissions, context, and execution environment. 𝟰. 𝗠𝗲𝘁𝗮-𝗛𝗮𝗿𝗻𝗲𝘀𝘀 → Governs multiple agent environments Coordinates different agents, tools, policies, permissions, and workflows. 𝗧𝗵𝗲 𝗺𝗲𝗻𝘁𝗮𝗹 𝗺𝗼𝗱𝗲𝗹: Loop = verification Graph = workflow Harness = capability Meta-harness = governance A better model can improve reasoning. But reliable agents depend on the architecture built around the model.

𝗧𝗵𝗲 𝟰 𝗟𝗮𝘆𝗲𝗿𝘀 𝗼𝗳 𝗮𝗻 𝗔𝗴𝗲𝗻𝘁 𝗦𝘆𝘀𝘁𝗲𝗺 An agent can burn tokens, declare success, and still fail the tests. That’s often an architecture problem, not a prompting problem. 𝟭. 𝗟𝗼𝗼𝗽 → Decides whether to continue Acts → verifies → retries until there’s evidence of success. 𝟮. 𝗚𝗿𝗮𝗽𝗵 → Decides what runs next Handles branches, retries, handoffs, fallbacks, and shared state. 𝟯. 𝗛𝗮𝗿𝗻𝗲𝘀𝘀 → Gives the model its capabilities Tools, APIs, files, memory, permissions, context, and execution environment. 𝟰. 𝗠𝗲𝘁𝗮-𝗛𝗮𝗿𝗻𝗲𝘀𝘀 → Governs multiple agent environments Coordinates different agents, tools, policies, permissions, and workflows. 𝗧𝗵𝗲 𝗺𝗲𝗻𝘁𝗮𝗹 𝗺𝗼𝗱𝗲𝗹: Loop = verification Graph = workflow Harness = capability Meta-harness = governance A better model can improve reasoning. But reliable agents depend on the architecture built around the model.

Dhairya Karekar

18,724 просмотров • 29 дней назад

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.

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

441,974 просмотров • 3 месяцев назад

this AI engineering breakdown from IBM Technology is pure f*cking treasure it explains the role as a connected system 01 Python and APIs give you the surface to build on: services, data, integrations, and the software around the model 02 RAG and embeddings solve the context layer: finding the right information before the model has to reason about it 03 agents turn that context into work: they choose tools, run steps, and move toward an outcome 04 observability closes the feedback loop: traces, metrics, failures, latency, and the evidence needed to improve the system 05 deployment is where the system becomes real: reliability, constraints, and the conditions that let people actually depend on it the model sits inside this architecture the AI engineer connects every layer into one operating loop: build → retrieve → act → observe → ship save this, then build your solo business with an AI employee ⭣

this AI engineering breakdown from IBM Technology is pure f*cking treasure it explains the role as a connected system 01 Python and APIs give you the surface to build on: services, data, integrations, and the software around the model 02 RAG and embeddings solve the context layer: finding the right information before the model has to reason about it 03 agents turn that context into work: they choose tools, run steps, and move toward an outcome 04 observability closes the feedback loop: traces, metrics, failures, latency, and the evidence needed to improve the system 05 deployment is where the system becomes real: reliability, constraints, and the conditions that let people actually depend on it the model sits inside this architecture the AI engineer connects every layer into one operating loop: build → retrieve → act → observe → ship save this, then build your solo business with an AI employee ⭣

beamnxw ./

13,505 просмотров • 16 дней назад

You've probably scrolled past a dozen posts about Jev this week without anyone telling you what it actually is. It's the first model from TypeSafe, a lab started by one of the researchers behind ChatGPT. It's a decision engine: you give it options, it picks one and tells you how sure it is. It cannot write a single word, and that is the interesting part. Every other AI you use writes. That is the whole interface. So when software needs a plain yes or no, we make a model write a paragraph and then dig the answer back out of it. Fine in a chat window where a human reads it. Bad inside software, where code has to act on it. The bet is that the valuable half was never the writing. It was the deciding. That problem showed up in WordPress years ago, and it is the reason WPVibe works the way it does. The AI does the work. Anything permanent stops and waits, because a delete that skips the trash is not something software should decide on its own.

You've probably scrolled past a dozen posts about Jev this week without anyone telling you what it actually is. It's the first model from TypeSafe, a lab started by one of the researchers behind ChatGPT. It's a decision engine: you give it options, it picks one and tells you how sure it is. It cannot write a single word, and that is the interesting part. Every other AI you use writes. That is the whole interface. So when software needs a plain yes or no, we make a model write a paragraph and then dig the answer back out of it. Fine in a chat window where a human reads it. Bad inside software, where code has to act on it. The bet is that the valuable half was never the writing. It was the deciding. That problem showed up in WordPress years ago, and it is the reason WPVibe works the way it does. The AI does the work. Anything permanent stops and waits, because a delete that skips the trash is not something software should decide on its own.

John Turner

21,029 просмотров • 17 дней назад

Seems like Visual Studio Code is starting to tell you: your agent primitives need to move. There is now a new migration banner in the Chat panel, and it is part of a much bigger change happening under the hood: the move from the old Local harness to the new Agent Host architecture built around AHP. This is not just about moving where an agent runs. The old model was very VS Code-centric: prompts, custom agents, instructions and skills could live in VS Code-specific locations and the agent runtime lived inside the extension host. The new Agent Host separates the agent runtime from the editor. Sessions can keep running when the window closes, be shared across VS Code windows, run remotely, and support different harnesses such as Copilot, CLI and Copilot Desktop App through a common session layer. And that means some of our primitives need to move too. Prompt files are being deprecated for Agent Host and migrated to Skills. User-level agents and instructions that lived in VS Code profile storage need to move to harness-supported locations. Even the old location settings are being deprecated. The new migration experience can detect these things and guide you through moving or converting them, while keeping the originals unless you explicitly remove them. The new banner is basically the first visible sign that this migration is becoming a real product workflow. Basically telling us - it's time to move on!!!! If you have accumulated a lot of prompts, custom agents, instructions and skills over the last year, now is probably a good time to understand where they actually live and which harness owns them. To summarize the shift - it isn't just: VS Code Chat → Agent Host It is: VS Code-specific primitives → harness-native primitives. And I think this is going to become increasingly important as agents stop being features inside an IDE and become runtimes that multiple clients can connect to. Go run your migrations now 🏃‍♀️

Oren Melamed

29,753 просмотров • 4 дней назад