I went a little overboard with Codex last week... and burned through my entire weekly allowance in two days. Luckily, my quota reset today. Otherwise, I’m not sure what I would’ve done. It got me thinking: instead of asking one large model to handle everything from start to finish, why not let a stronger model plan the project and review the work, while a model built for execution handles the day-to-day implementation? So I tried it. The result was better than I expected. I used GPT-5.6 Sol in Codex as the decision-maker, then ran Ling-3.0-flash from Ant Ling inside OpenCode as the execution engine. Together, they built a small 3D farming game. Before writing any code, I had Codex create four documents: SPEC.md defined the product scope and the lines we couldn’t cross. ARCHITECTURE.md laid out the isometric coordinate system, state machine, and module boundaries. TASKS.md broke the project into small jobs Ling could tackle one at a time. ACCEPTANCE.md explained how each step would be tested and what “done” actually meant. Then I gave Ling a very straightforward role: You are the execution model for this project. Read all four documents before you begin. Work only on the task assigned for this round. When you’re done, run typecheck, test, and build. If anything fails, read the error, fix it, and run the checks again. Do not move on to the next task early. Ling handled dependency installation, project structure, strict TypeScript configuration, test setup, and a production build in 6 minutes and 3 seconds. It ran into issues with the Vite test config, a TS6310 error, and a missing jsdom dependency along the way. Instead of stopping at the first error, it kept reading the logs and fixing the problems until all three checks passed. The speed was honestly hard to believe. If you exclude the time spent waiting on tools, it was producing more than 100 tokens per second. That made the whole development loop feel noticeably faster. After this experiment, I’m planning to keep using the same workflow. If the task is small, there’s no reason to call an expensive planning model for every single step. If the task is large, handing the entire project to a Flash model in one prompt isn’t a great idea either. The setup that makes more sense to me is: Use a more capable model such as Codex to explore the project, make architectural decisions, and break the work down. Put the constraints into specs, schemas, types, and tests instead of leaving them buried in chat history. Give Ling-3.0-flash a steady stream of clear, verifiable implementation tasks. Report bugs with structured context and actual error logs, rather than saying, “It still doesn’t work.” Bring Codex back in for architecture reviews, visual checks, and changes that affect multiple parts of the project. The point of this setup isn’t to give AI a big “build the whole project” button. It’s to turn software development into a pipeline with a much more sensible cost structure: Codex figures out the plan, sets the boundaries, and catches problems. Ling-3.0-flash moves quickly, calls tools reliably, and works through well-defined tasks at scale. For agent workflows that involve lots of repetitive edits, production tasks, and tool calls, this may be a more practical answer than simply using the biggest model for everything.show more

this is the best trick to maximum usage limits... on chatgpt codex codex's best kept secret is that your main agent doesn't have to do everything... custom agents are just files in ~/.codex/agents, and one file gives you a second worker on deepseek v4 flash > create ~/.codex/agents/deepseek-worker.toml > set model = "opencode-go/deepseek-v4-flash" with model_reasoning_effort = "max" > keep it bounded: one task packet, no scope creep, report back ```toml name = "deepseek_worker" description = "bounded implementation, testing, and cleanup on deepseek v4 flash" model = "opencode-go/deepseek-v4-flash" model_reasoning_effort = "max" ``` then @ deepseek_worker in the composer... your root agent plans while the worker ships the implementation planning on the main model, execution on the flash lane... that's the whole trick (we run this exact file, last i checked it keeps the heavy turns off the main thread)show more

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 ↓show more

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 ↓show more

2 Cursor Composer agents working together on the same... project. one builds the project, the other reviews once the first agent is done and writes a report. cursor rules file is in comment to achieve this. first agent is instructed to build the project second agent is instructed to write a monitoring script to determine when the project is done then to review the project for errors and write a report This prompt can be improved. This is meant as a baseline If you are interested in learning how to use Cursor more in depth, I have a 28 chapter 1000x Cursor Course building full apps from scratch with 19 hours of content. link is in my bio if you are interestedshow more

This is my "feel the AGI" moment: I used... GPT-5.6 Sol to train my own autocorrect model that outperforms GPT-5.6 Sol (wtf??) I have no ML background. I have no idea what I'm doing. I just kept pushing Sol until it spat out a SOTA model. And I spent $0. The motivation: Years of talking to AI have made me terrible at typing. Rather than fix my skill issue, I decided to throw more AI at it. My idea was: instead of autocorrect that interrupts my flow, I want to type fast with mistakes and have AI clean it up after. I wanted the smallest local model possible, for speed, for battery life, for science! So I decided to train my own. Inspired by Andrej Karpathy’s autoresearch, I ran Codex /goal with this setup: pick an experiment, try it, record the results to a doc, throw it out if it fails, and plan the next experiment without repeating failures. I gave a few examples that had to pass, tight latency targets, and let it run. Sol did some amazing things. First, it scanned benchmarks and shortlisted base models: Qwen 3.5, Gemma 4, Liquid LFM 2.5. It found a dataset on HuggingFace for typed text. Then it built a simulator for fingers striking a Mac keyboard, modeling the physical layout with a Gaussian distribution around each key. It simulated striking the wrong key, wrong order, fat-fingering, etc. With the models + data + simulator, it fine-tuned using MLX right on my MacBook. It had a working prototype within an hour! But accuracy was pretty poor. — Problem 1: Tokenization Sol read papers, ran tests, and identified that the tokenizer was the bottleneck. Tokenization makes typos hard for the model to see, so it memorizes mappings instead of using its language priors. Sol tried ByT5, Google’s tokenizer-free byte-level LLM. This made a big improvement, but the model is old and lacked the knowledge needed to reach Sol performance. Sol dug deeper and realized a tokenizer-free model isn’t needed; instead, it used T5Gemma, an encoder-decoder model. This can understand the input deeply before producing output, and furthermore, Sol could post-train the encoder to improve performance. This gave a much higher ceiling. — Problem 2: Loss function Now the model was correcting some typos perfectly, but ignoring most. Sol realized that standard cross-entropy loss was teaching the model to avoid edits, because the vast majority of characters in the training data were left unmodified. The fix was wild: Sol wrote a custom loss function that byte-aligns the source and target strings, uses a dynamic programming algorithm to compute the minimum edits between the two, then weights correct edits much higher than copies. After a lot of tuning, this dramatically improved accuracy. — Problem 3: Autoregression One failure mode remained: if the model made a mistake, it couldn’t backtrack. It could only predict the next token. Teaching it to “think” like a reasoning model would solve this, but would be far too slow. Sol found a beautiful solution: instead of greedily predicting the next token, beam search over all possibilities. This parallelizes the exploration instead of one linear chain-of-thought. At the end, choose the path with highest cumulative log probability. This worked great, but made the experience worse, since the user wouldn’t see progress until the whole search was done. To fix this, Sol made a clever observation: after each search step, the longest common prefix among surviving branches is guaranteed to appear in the final result, so it can be displayed immediately. As the search progresses, weaker paths are dropped and the prefix grows, so the user sees continuous progress. Sol built all this as a custom MLX pipeline that does the parallel decoding on the MacBook GPU, with just ~40ms TTFT. It’s crazy fast and entirely local. — Final eval (error reduction rate, higher is better): - Apple autocorrect: 49.66% - GPT-5.6 Luna: 82.47% - GPT-5.6 Terra: 87.64% - GPT-5.6 Sol: 90.56% - Our model (1.7B): 91.02% Final cost: - 1 quota reset (thanks Tibo) - $0 (And yes, I verified there's no cheating. In fact, we test words scrubbed from the training data to prove the model isn’t memorizing) There were a ton more details and tangents I could write about: contrastive learning, GRPO, DPO, dynamic masking, and more. Sol is a fascinating and creative model. It blew my mind so many times. Don’t let a lack of experience stop you: Sol makes AI experiments accessible to anyone!show more

i just open sourced the workflow behind $2M AI... video productions... i built 7 skills that run the pipeline end to end, built for Seedance 2.5 and they work in Claude Code, Codex, Hermes or any harness (works best with 1080p using Higgsfield CLI) here's how to use them, in order: /setup writes which image and video models you run into your project, once, so every skill reads the same stack /studio-init scaffolds the whole studio as a file tree from one question, the project name /film-breakdown walks your script scene by scene and writes a 22-field card for every shot /reference-board locks your references into a visual bible, a caption on every image and a ban list for the rest /asset-passport writes the exhaustive descriptor every later prompt will quote word for word /stress-test combat-tests each asset and flips it to locked only at 10 out of 10 repeatability /shot-prompt refuses to run until everything in frame is locked, then writes the 15-block prompt and logs every attempt get access to the skills and full breakdown of the pipeline in the article below:show more

I’m joining OpenAI Codex to work on the future... of agentic development! At Cursor, I got to see the shift from autocomplete to agents. The next step isn’t a better IDE. It’s an Agent Development Environment (ADE): systems and tools for orchestrating agents, reasoning over their outputs, and making them autonomous enough to reliably complete ambitious work. After chatting with Alexander Embiricos and Tibo, it was clear that Codex is the best place to realize this vision. The team has consistently shipped SOTA models for agentic coding (check out gpt-5.3-codex) and I’m pumped for the future that the new Codex App points to. What I’m most excited about is the broader mission: accelerating the knowledge work economy. All agents are coding agents, and we’re already seeing Codex used across every job function within organizations. I’m extremely grateful for my time at Cursor, working with the incredible team, and I’m proud of what we built together. I’m excited to take an even bigger swing with Codex. If you’re curious to get a glimpse of where we are headed, download the Codex App! If you want to work on this mission, please apply or reach out - we are hiring across all functions! You can just build things.show more

🚨Gemini 3.6 Flash is trash I tested it on... a 3D Golden Gate Bridge, and the results were awful. • I had to re-prompt it three times because it repeatedly ignored the instructions. • First attempt, instead of creating the requested .html file, it first tried to build the experience inside the Gemini app using simulations. • Then second attempt it started placing images from the web into the chat rather than actually producing the file. • Even after getting it to complete the task, the final output was dramatically worse than Gemini 3.1 Pro, which is 5 months old and now not even a top 10 model on leaderboards. This feels like a regression from Gemini 3.5 Flash and honestly, it is one of the weakest models I have tested in the past few months. Has anyone else tested Gemini 3.6 Flash yet, and are you seeing the same thing?show more

You can now orchestrate Fable 5, Sol, and any... model inside Codex with one plugin. It's called Codex-Orchestration. Assign Fable 5 as the advisor, Sol as the executor, or any model to any role. Then define the order they work in. Codex handles the routing. I ran Fable 5 High as planner with GPT-5.6 Sol Extra High as executor on a set of issues Opus and GPT-5.5 always struggled with. Done in 30 minutes. 40% fewer limit hits. 2x faster implementation. Install it by pasting this into Codex: "Install Codex Orchestration: codex plugin marketplace add Cjbuilds/Codex-Orchestration codex plugin add codex-orchestration@codex-orchestration Verify the installation, then tell me to start a new task." Then assign your models: @ codex-orchestration advisor: Claude Fable 5 High, Executor: GPT-5.6 Sol High Open source. Tweak the routing however you want.show more

Robotics keeps hitting the same wall. Single task RL... works, but... it does not scale to hundreds of tasks or new embodiments. This new paper looks like a real step toward fixing that. The team introduces MMBench, a benchmark with 200 tasks across many domains and robots, and Newt, a language conditioned world model trained online across all 200 tasks at once. The simple idea behind Newt: The model learns from demos to get the right priors It trains across many tasks through online interaction It uses language to ground the goal It adapts fast when a new task shows up What stood out to me: ✅ One model trained on 200 tasks at the same time ✅ Language conditioned control for both states and RGB ✅ Better data efficiency than strong baselines ✅ Strong open loop control ✅ Fast adaptation to new tasks and embodiments ✅ Full release of 200 checkpoints, 4000 demos, code, and benchmark This is a good push toward general control instead of one model per task. If you want the full paper: Project page: —- Weekly robotics and AI insights. Subscribe free:show more

People still don't understand how good Codex has become... 👀 I asked it to build this Starbucks website and within some 30 minutes it generated this without me touching a single line of code or spending time looking for assets either. Imagen generated every visual and GPT hooked everything into the codebase. After a couple of small tweaks it all just worked. It didn't even use half of my Codex session. It's super surprising to see quality of the assets. Was way better than I expected. That made me think about what AI is actually changing. Being a developer or a designer was never just about doing the work, writing code or designing it using some kind of tool. It was about knowing what to build in the first place. AI is getting incredibly good at the doing. The hard part is still having good ideas, making the right decisions, and knowing when something feels right. That is only getting more valuable. The people who win won't be the ones writing the most code. They'll be the ones with the best ideas who know how to turn them into reality with AI. Live: Code:show more

▣ Introducing Endless: infinite inference (kinda). An experimental harness... to milk every ounce out of your Codex subscription. Since Codex can let an in-progress turn keep going even after your usage hits 100%, why not put that to the test? Endless starts one Codex turn and gives the agent a wait_for_user_input tool. Once it finishes a task, it calls that tool and waits. Your next message becomes the tool result, keeping the entire session inside the same turn. It runs through Codex’s own app server using your existing ChatGPT login. Native tools, automatic compaction, context tracking, and quota tracking still work as usual. ⚠️ NOTE: I CAN’T CONFIRM THAT YOU WON’T GET BANNED OR PUNISHED FOR USING THIS TOOL. USE IT AT YOUR OWN RISK.show more

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.show more

Ling-3.0-flash is built as a fast, reliable execution engine... for agent workflows. It shines in long-running tasks, tool calling, and high-volume production work where speed and stability matter more than massive reasoning depth. At 124B parameters with only 5.1B active, it keeps costs low while delivering quick responses and strong instruction following I ran a practical task using Ling-3.0-flash capabilities (based on its documented agent strengths in coding and tool use). Real Test Run I tested Ling-3.0-flash on a content creation task that matches your workflow. I asked it to generate a short, viral-style YouTube Shorts script for football highlights, including captions, title suggestions, and thumbnail ideas. Input Prompt: "Create a 30-second YouTube Shorts script for a dramatic Ronaldo goal from the 2026 World Cup qualifiers. Include engaging English narration, 3 multilingual caption versions (English, Spanish, Portuguese), a catchy title, and thumbnail description. Make it feel real and exciting for football fans." Process: The model first outlined the structure: intro hook, key action description, emotional peak, and call to action. It generated the script, then created caption variants, optimized the title for clicks, and suggested a thumbnail layout. It handled iterations well when I asked for adjustments, such as making it more dramatic or adding player stats. Total interaction took under 2 minutes with low token use. Result: - Script: Solid, ready-to-record narration with natural flow. - Captions: High-quality and culturally adapted. - Title & Thumbnail: Click-worthy and on brand. This demonstrates its strength for creators who need fast, high-quality content assets. Official account : Demo:show more

After a few more hours, I think I've figured... out Opus 5. Opus 5 is trained to be more agentic than anything I've used. All Claude 5 models are like that. So what changes? The way to interact with Opus 5 or contextualize it won't work the same way as with other models. It loves exploring, so it doesn't need much guidance for it. Unique preferences, artifacts, and references compliment it well and enable cleaner and more effective exploration and execution. Now that it can explore more effectively on its own and understand intent better, the best thing to do is to get out of its way (e.g., it doesn't need examples of your preferences; a clear high-level description of it works best). It's truly agentic in that sense. A good first step to provide better context for Opus 5 is to distinguish between what's situational and what needs persistence. Regardless, persistent system prompts and CLAUDE.MD needs to stay lightweight. Remove memories and tool descriptions from these. CLAUDE.MD is also a great place to tap into progressive disclosure by linking command/skills to it. On the situational side, agent skills and auto-memory can leverage progressive disclosure and the improved ability of the model to use its external context/knowledge. Conflicting and unnecessary instructions, which are common at this layer (mainly to ensure reliability), are going to throw off this model easily. That's the biggest change I had to make. Simple, clean, and clear prompts and skills work best. I had to clean a lot of my skills and system prompts. The way I prompt remains the same (usually clear and well-scoped). MCP tool descriptions are also more descriptive and have been deduped from the system prompt. Anthropic released a guide on the new rules for context engineering, which was helpful here. I started to test the recommendations and created a little artifact with the things that worked along the way. This might feel like a lot of work. Believe me, it has been frustrating. But I think we can expect future frontier models to become more agentic and smarter at figuring out the right context/gaps. The best thing to do is to prepare for that now. Boris Cherny mentioned that Opus 5 is their least prompt-injectable model yet. I am not sure if that was something they intentionally trained for or if it emerged based on how it was trained, which is to be extremely agentic in nature and more direct in execution.show more

Simplicity is at the heart of great software. This... is one of the reasons why Claude Code has been sticky for me. As a builder, I love planning and brainstorming, and this is now a key focus of Claude Code. I use Shift + Tab a lot to cycle between brainstorming, planning, and execution. This functionality provides the appropriate interface for me to either be very involved or less involved as I please. This works particularly well when building out new and complex features or entire new projects. This saves a huge amount of time. It allows me to tune Claude Code to execute and build more effectively. It also builds a loop of trust, and I often (surprisingly) find Claude Code asking for clarifications when it's confused. Coding agents don't normally do that. I have shared before on the power of brainstorming with AI for longer times. Try it and you will not be disappointed. Vibe coding is fun, but pair it with intentional development cycles, and you watch how far you can take a project with coding agents today.show more

I am very proud of my work at Royal... Caribbean, and I was honored to be set the complex task of reinventing their entire entertainment product from the ground up (by Lisa Lutoff-Perlo and Michael Bayley), and to make it the very best in the industry, if not the world. This lofty mission resulted in an incredibly rewarding decade for all involved. Massive changes delivered massive results. Job done👍. So, as a way to recognize the incredible support that the brand gave to all of my ideas, concepts, developments and inventions (and as a fun way to give back), I always included the famous “Crown & Anchor” logo in every entertainment production. Sometimes I did this openly and loudly and sometimes more subtly, as an “easter egg” for people to discover. It’s always there tho. (Have you spotted all of them?) Here are three examples of my not-so-top-secret “Show The Logo” strategy. – An example of a small hidden one, a medium-sized subtle one, and a LOUD ‘n’ PROUD HUGE ONE, that hides from nobody. Happy hunting. 😉show more

Codex can run Qwen-3.8-max now as well!! Alibaba most... capable model, dropped today and it's already in my codex picker. It's a token plan subscription, not metered api billing. You take the key from your Qwen plan, drop it into Codex Router, and it spends down the plan instead of your card. There's a catch though. Qwen's official setup switches your whole codex over to them, so your ChatGPT models stop showing up at all. That's exactly what Codex Router is for. It adds models to the list instead of replacing them, so sol, Grok, kimi, Deepseek and now Qwen 3.8 max all sit in the same picker and it can grab whichever one suits the job. Router's open source, setup's in the video 👇show more

whoever leaked this has bigger balls than sense someone... gave a fleet of Claude agents shared memory so they would stop contradicting each other, then measured both the bill and the output: the version that talked most made 2.4x the api calls of the version that won, and hallucinated 34% more than doing nothing at all, 0.658 against 0.492 i ran the same question past two of my own agents afterwards and got two different answers about which file owns the config. each one was individually right and the pair was wrong, which is the whole failure in one line this is Graph Engineering, the layer that decides which agents may talk to each other at all, and it installs into the agent you already pay for: - decide which agents may share state at all, because every edge you draw is a channel a mistake can travel down - measure divergence per PAIR instead of as a fleet average, across what they believe about place, time and task history - gate on that number and stop the pair above your threshold before it reasons, rather than repairing the output afterwards - let compressed summaries replace whole states: the verified protocol landed 0.463 against 0.658 for full broadcast - cut the sync frequency until it hurts, since the winning setup used 58% fewer calls than the one that broke it - never propagate a state nobody checked, because the contamination effect came in at d=1.18, a full standard deviation of extra lying - keep the shared layer small enough to diff, which is what a written standard does and a running conversation cannot - re-run the check after every model upgrade, because this was 8 scenarios on one model family at n=30 per condition - and learn where it does not bite: on plain software tasks every condition converged under 0.2 and the whole effect vanished turns out the ranking is the uncomfortable part: verified summaries 0.463, no synchronisation at all 0.492, full broadcast 0.658. the middle option is doing nothing, and it beat the thing everyone builds first the group agreeing is what it looks like when every agent copied the same mistake, which is why a fleet that hallucinates has a replication problem and keeps getting handed a smarter model instead so the question for your own setup: if you asked two of your agents the same thing right now, would they answer the same way bookmark this one. the layer underneath it, deciding which arrows between agents exist at all, is built step by step in the piece below ↓show more