Over two weeks, Prime Agent orchestrated a swarm of over 2,000 agents to rewrite itself in Rust. With 10,000+ sandboxes, 200B+ GLM-5.3 tokens, and 16,000 agent-to-agent messages, this was our largest test yet of Prime Agent’s multi-agent capabilities and our infrastructure powering these agent swarms. Prime Agent now reaches usable input ~13× faster and uses 83% less startup memory than before.

I remember when at my previous company Doctrine we rewrote our frontend from AngularJS 1.7 to React. It was a huge project and a big decision. It took roughly six months and a team of talented software engineers. That’s why the Bun rewrite is so interesting to me. Bun was ~535k lines of Zig. The rewrite produced more than 1M lines of Rust diff! But... it took only 11 days. At peak they had around 64 agents running in parallel, across 4 worktrees, with 16 agents per workflow. It generated 6,500+ commits (wow!) It was also a lot of work because they didn't ask "plz Code go from Zig to Rust" but they created a porting guide, a flow to analyze errors, separate adversarial agents to review diff to find mistakes etc. The human job was not to write the code but more designing and supervising the process aka "managing agents." IMO it's a big deal. The world is full of important software stuck on old frameworks, old languages, old architectures and old security models because rewriting is too expensive and too risky. If this methodology becomes reliable, we can move software to safer languages, better runtimes, cleaner architectures and more secure defaults much faster. So much software run legacy COBOL and soon the transition to Rust will be easy.

5.5 is smart enough to cram everything into a single agents.md file with all your repos in a single workspace. Containerize and break everything into microservices with TDD for everything with its own repo and pipeline. Put docs in html in their own repo and enforce it in HTML with hyperlinks so you can read them as well easily. By splitting out the services more you can have agents recursively spawning subagents by devloop and splitting out work as effectively as needed on worktrees or patches in each repo. I use apfs/git worktrees are setup by default in omp for subagents. 5.5 is smart enough to know exactly how many subagents it needs and what depth the agent tree should go with recursion turned on. You can refactor, rewrite, test, deploy your entire codebase with a single prompt. I'm currently managing a ~6m line codebase this way and self hosting everything. All infra code has its own repo for helm/csps. Metrics and tests on all services and grafana dashboards for tracking everything. Every stat I can possibly think of or ever will need gets shoved into postgres. Any heavy data work or code that needs to be written gets metrics assigned to it in prom or pg and then gets targeted for /goal or for /autoresearch for optimization. I typically explore everything with gemini and have it write the prompts to hand off to 5.5. People are simply not pushing the models hard enough. 5.4 and 5.5 (maybe even kimi and ds) ARE ASI. The bottleneck is yourself, compute, and devloop times.

The biggest upheaval to work in our lifetime is here. Joe Hudson (@FU_joehudson) coaches the research team at @OpenAI (along with @sama and leaders across Apple and Google), and from that front-row seat, has noticed the people doing best in these AI-forward environments aren't winning for the reasons you'd expect. It's not more knowledge or more effort. Those are the exact two things AI does best. Instead, they win on "emotional clarity": staying in hard conversations, not turning on themselves (and others) when things get tough, and pushing forward through failure. In today's incredible guest post, Joe shows you exactly how to build this skill within yourself and your team. Inside: + How AI is changing work: the NBA-ification of teams + The new inner game: the "wisdom stack" that will set you apart + How to build a team that can metabolize fear Don't miss this one: lennysnewsletter.com/p/the-n…

This problem will naturally tend to go away as companies are grown from the start using AI. Then you don't need to extract any domain knowledge from people's heads; it will never have been in people's heads.

OpenAI and Anthropic are effectively telling the market they can't solve every problem with a generic AI coworker. You don't pour billions into massive forward-deployed joint ventures if you think the next model release is going to take care of it. In the cloud supercycle, semis led and software followed (and you didn't need Qualcomm or ARM to tell you the value was migrating up the stack). In AI, the infra layer itself is telling us the application layer is a separate, massive opportunity they can't fully capture. a16z's @joeschmidtiv on why the app layer isn't dead: a16z.news/p/avoiding-death-o…

The Polsia public dashboard sits at archive.ph/S5uq2 and Claude and I spent 15 minutes reading the JSON so you don't have to. What follows is what's actually on it, in plain language, with the meaning of each number stated alongside the number itself. The headline figure is 5,010 "companies," and that word is doing more work than people realize. These are not 5,010 real businesses with paying customers and revenue lines. They are 5,010 instances of the Polsia software, each one a user account where someone spun up what the platform calls an AI operator, and the dashboard records what every one of those operators produces. Almost none of them produce anything at all. Paid churn is 63.5 percent in 30 days, which means roughly two out of every three people who handed over money for the platform a month ago have already walked away from it. Healthy SaaS churn at this stage of company life runs in the single digits, which makes this number roughly ten times worse than the floor of what a venture-grade business should be losing every month. ARR is shrinking by 39 percent week over week, which is a sentence worth re-reading. The company that just announced a $30 million Series A is watching its annualized revenue line go down, meaningfully, every seven days. Daily inference cost is $27,272, which is the spend on AI model calls keeping all 5,010 of those operators alive and producing their CEO reports every day. The cost is real, it is burning right now, and the output of that burn is the paragraph below. Every operator CEO report visible in the snapshot reads the same way: zero customers, zero revenue, no shipped product, and then the AI writes an optimistic plan for tomorrow underneath the zero-traction admission. That optimism layer, generated on top of nothing, is what the platform is selling its users. In plain language: a founder built an app that runs LLM calls in a loop to generate CEO reports for businesses with no customers and no revenue, charged users to participate in the loop, announced a $30 million round while the underlying business burns cash and loses paying users faster than it gains them, and described the raise as one his AI ran for him (see comments to read what Claude wrote about this). The dashboard is public and he chose to leave it public, which means the receipts have been sitting on his own infrastructure the whole time. I am not a journalist and I am not auditioning to be one. I am a software engineer who reads JSON and uses AI to decipher it quickly, and in this case the JSON is the JSON.

We have, as far as I can tell, no good tests of the productivity impact of the autonomous coding tools that appeared starting in December 2025. Every paper out there is from prior to the Claude Code/Codex revolution. A huge gap in our knowledge about what is happening in coding.

1/ Some things I've learned recently running coding agents on large-scale projects. Most of this contradicts advice from 6 months ago!