Autonomy is compounding on both sides of the stack today: a research agent rewrote its own tooling in eight days, OpenAI trained a system whose only job is attacking its own models, and investors sold AI stocks through one of the cleanest earnings beats the sector can produce.
The Ghost in the Machine: Unsupervised Autonomy
Four projects this week hand agents more control over their own operation, from rewriting code to spinning up and tearing down their own environments.
- An AI Research Agent Rewrote Its Own Harness in Eight Days. Weco AI’s self-tuning system outperformed code its own engineers spent two years hand-tuning, though the team stops short of calling it self-sustaining.
- Goodfire Opens Beta for an AI Agent That Runs Research Experiments. Silico, billed as an on-demand research team, reproduced a monthslong interpretability project in two days, according to its maker.
- A solo coder built a $110 pipeline that merges its own PRs. Andy Widjaja’s autoloop system triages, breaks down, and ships backlog work on a Claude Max subscription and a $10 monthly server.
- Perplexity ships SPACE, sandboxes that self destruct after each agent task. Firecracker microVMs, outside credential handoffs, and rolling snapshots now anchor the Computer agent platform, wiping each sandbox once a task ends.
New Frontier: Models Built to Judge and Attack Themselves
The week’s model news is less about raw scale and more about systems trained to critique or attack their own kind.
- OpenAI built an AI that breaks its own models to make them safer. GPT-Red, an automated attacker OpenAI trained in house, found exploits that made GPT-5.6 six times harder to hijack.
- GPT-5.6 Sol Jumps 18 Spots to Top a Web Design Leaderboard. OpenAI’s model climbed to first place by learning to spot the generic patterns AI built sites fall into, then refusing to reuse them.
- Thinking Machines Ships Its First Open Model, a 975B MoE. Inkling pairs a one million token context window with 41 billion active parameters and can be tuned through the company’s Tinker toolkit.
- Nvidia’s Cosmos 3 Edge lands as Fujitsu, Hitachi join Japan push. Nvidia’s edge world model is shipping for robots and factories as Japan commits 27,500 Rubin chips to a new industrial coalition.
The Coding Agent Wars: Open Code, Closed Process
Coding agents keep shipping, but the fine print over openness, cost, and testing is where the real differences show up.
- xAI Opens Grok Build’s Code, Not Its Development Process. The coding agent’s source went public on GitHub, but the license blocks outside pull requests and mirrors a private monorepo rather than an open project.
- ReactBench Puts Coding Agents Through Real Pull Requests, Not Puzzles. OpenAI’s top configuration cleared 43.1 percent of React tasks on the first try, while Anthropic’s best challenger cost six times more per run for a lower score.
- Open Interpreter’s GitHub Repo Turns Coding Agents Into UI Testers. The open source project lets an agent write and run code locally, then operate native or web app interfaces to confirm the result actually works.
- OpenAI’s Supply Co. sells a $230 keyboard built to run Codex. The kbd-1.0-codex-micro ships with dedicated Codex keys, a reasoning-effort dial, and a workflow joystick, and it is already listed as out of stock.
Money, Chips, and Skepticism: The Business Underneath the Boom
Behind the product news, the financial and physical infrastructure of AI keeps testing how much confidence the market still has left.
- Anthropic’s Bankers Start Booking Investor Meetings for IPO. JPMorgan Chase, Morgan Stanley, and Goldman Sachs are lining up investor meetings, moving Anthropic’s potential listing from paperwork into an active pitch.
- TSMC beat and raised. The market sold AI stocks anyway.. Traders cut 1.1 percent off the Nasdaq 100 on Thursday, deciding TSMC’s strong guidance no longer justifies fresh gains for the AI trade.
- IBM Research: routing AI tasks by price tag alone gets it wrong. New benchmarks from IBM Research find that caching and serving infrastructure shape AI costs more than a model’s sticker price does.
- NVIDIA shrinks its robot brain: new Jetson Thor chips target scale. The T3000 and T2000 modules match T5000-level performance in half the size, with Amazon Robotics, Boston Dynamics, FANUC, and 1X already building on the platform.
- Systolic Arrays Handle Most AI Chip Math, an Engineer Argues. Joe Fioti argues that the decades-old systolic array design, not raw transistor count, now determines how much of a chip’s peak throughput actually reaches real workloads.
Today’s Quick Hits
- Vercel Opens Its AI Gateway Leaderboard Data to the Public. The company licensed its model, lab, app, and provider usage rankings under Creative Commons and added CSV exports plus shareable chart images.