AI Builders Digest

Anthropic traces a month of Claude Code degradation to three separate changes

Three unrelated changes stacked up into what looked like broad model degradation. Claude Code's default reasoning effort quietly dropped from high to medium on March 4 and is now back at xhigh for Opus 4.7; a March 26 caching optimization meant to clear old thinking once instead cleared it every single turn, which explains both the forgetfulness and the usage limits draining faster; and an April 16 system prompt line capping responses at 100 words cost 3% on evals. All three are fixed as of v2.1.116, and usage limits are reset for every subscriber.

• #products
• #evals

With Google on board, open weights gets a complete industry endorsement

With Google now on board, Levie reads the current wave of support as a complete endorsement of open weights AI rather than a niche position. He calls it a big moment for the industry.

• #open-source

The US AI community converged on open weights in under a month

A month ago open-weight support was contested, and now it feels obvious. Madhu Guru credits a run of public experiments that each exposed something different about incentives, innovation, and geopolitics: DeepSeek, the Microsoft-OpenAI breakup, GLM, Kimi, Fable, and the OpenAI-Hugging Face episode. His larger point is methodological. In uncharted territory, answers to hard questions only come from repeated contact with reality, and that is how the field will settle most of AI's big societal questions over the next decade.

• #open-source
• #policy

Benedict Evans: the models all look the same, so value moves up the stack

Three to six companies are making frontier models and they are all roughly the same, which Evans reads as evidence that no real barrier to entry exists yet: a lab can flunk out completely and be back near the top of the leaderboard, and that is a bad signal for the labs, not a good one. He points at mobile networks, where traffic rose more than a thousandfold, capex runs $200B a year, and all the value went up the stack to Uber and YouTube. On job loss he is blunt about the star charts claiming a model can do 93% of a first year law associate's work: you cannot measure what the associate does, and you cannot measure whether the model can do it, so 'it's horseshit.' The real bottleneck is not capability but that most people are not tool builders, do not see the problem a tool would solve, and are not in a position to build it anyway.

• #markets
• #products

Rauch: the software factory is the product, not the app it builds

Rauch's framing is that your product is only as good as the agents you set up to autonomously maintain it, the same lesson he says Tesla learned about manufacturing. When a new idea shows up, the move is not ad-hoc prompting an agent but building the factory that will start, maintain, and grow it. He runs his own research this way already: a plain research/ folder, an AGENTS.md describing the format and best practices he likes, and an agent CLI that correlates knowledge across past sessions. No knowledge graph, no app, syncs over iCloud or git, and the agent can render and deploy an HTML report when he wants to share an outcome.

• #agents
• #products

Claude Managed Agents can now run tools inside your own perimeter

Managed Agents can now execute tools in a sandbox you control, either your own infrastructure or a managed provider like Cloudflare, Daytona, Modal, or Vercel. The agent loop stays on Anthropic's side while code execution, files, and services stay behind your network policies and audit logging. MCP tunnels, in research preview, let agents reach internal databases, private APIs, and ticketing systems through a lightweight gateway that makes a single outbound connection, with no inbound firewall rules or public endpoints. Self-hosted sandboxes are in public beta, and Amplitude, Clay, and Rogo are already building on the combination.

• #agents
• #products

Anthropic split the agent's brain from its hands to stop losing sessions

Putting the session, harness, and sandbox in one container made that container a pet: if it died the session died with it, and debugging meant shelling into a box that also held user data. Managed Agents now treats each as a separate interface, so a crashed harness reboots with wake(sessionId), replays the durable event log, and a dead sandbox surfaces as a tool-call error Claude can retry. Containers get provisioned only when a tool call actually needs one, which dropped p50 time-to-first-token roughly 60% and p95 over 90%. The security payoff is real too: credentials never sit in the sandbox where Claude's generated code runs, with git tokens wired into the local remote at init and MCP OAuth tokens held in a vault behind a proxy.

• #agents

ChatGPT Work has passed Codex in active users

Sottiaux says ChatGPT Work now has more active users than Codex. He is also pushing a voice capability already live in the ChatGPT app that, in his framing, finally makes talking to your computer do something: 'It was always possible to speak to your computer. It wouldn't do much in return. But we fixed that bug.' He calls it a game changer from mobile.

• #products

Running Codex all day with 12 subagents hunting 200 bugs

Steinberger has been doing massive parallel QA ahead of an OpenClaw release and posted the literal prompt: full end-to-end test with live API keys, 12 subagents splitting up functionality, dev gateways on different ports for stress testing, worktrees, autonomous PRs, a goal of 200 bugs, fix the root cause every time with no band-aids, and a running markdown test report. What changed is durability. This workflow used to fall apart at compaction boundaries or the model would start cheating, and now it holds together and finds complex behavior issues. Separately he notes that OpenAI signed the letter while Ant has stayed silent, alongside two points he thinks are worth stating plainly: competition is good for the ecosystem, and serving models at scale is hard.

• #agents
• #evals

If you can build a software factory, you can build a factory factory

Nan Yu pushes the factory idea one level up: a system that produces software implies a system that produces those systems. He also thinks the abstraction is not really about software. Anywhere there is a designed and implemented intention, whether that is public health or law, the same recursion applies.

• #agents

Masad's chess engine hits roughly 1200 Elo on one finetuned LLM

Masad deployed a new chess engine he estimates at around 1200 Elo, with the goal of 2000+ under two hard constraints: one small finetuned LLM, no custom pretraining or architecture, and moves produced by the model with zero chess engine assistance. He is explicit that relaxing either constraint makes the problem much easier.

• #evals

Only a founder with total control could buy an astrology app

Kothari's read on a generative media company acquiring an astrology app is that the deal is only possible under two conditions: the CEO has complete control, meaning profitable with no board and no VC he is aware of, and is as ambitious as David Holz. Try explaining that acquisition to a broader team at a normal company and it does not survive the meeting. He is curious rather than dismissive about where it goes.

• #funding