Anthropic’s second day of trading went the wrong way. The shares changed hands near $84 in Friday morning trading, 14% below Thursday’s $97.17 close and about a dollar above the $83 offer price, per 24/7 Wall St.
The agent builders had a busier day than the traders. Vercel opened a shared task queue for coding agents, Perplexity let its agent write its own playbooks, and OpenAI explained how Codex agents chase ChatGPT slowdowns.
Two companies also graded coding agents. Replit says 32 Claude agents sped up NumPy 18.9 times, and Stripe says agents pass its strict bar only 29.4 percent of the time, on a test that scores its own plugin.
The rest of the issue covers Chrome’s fake internet for attack testing, DeepMind’s four roads to superintelligence, and a Gemini feature that clicks through your Mac.
The Premium Evaporates: Anthropic Meets the Market
A debut that opened at $111.50 has already handed back almost everything. The question now is whether $83 holds as a floor.
- Anthropic stock falls to near its $83 IPO price on day two. After a $97.17 first close, shares traded 14% lower in Friday morning trading. 24/7 Wall St. points to a rotation out of AI growth stocks, while CoreWeave fell 6% and Amazon slipped 2%.
Agents Get a Manager: Queues, Playbooks and Fleets
Four companies describe what happens once one coding agent becomes dozens. Most of the figures come from the companies themselves.
- Vercel Board gives fleets of coding agents a shared task queue. Agents from different vendors can claim work, log progress and hand results to humans. Vercel says its v0 pipeline now merges more than half its pull requests this way, with no baseline given.
- Perplexity lets its agent write its own playbooks from your past work. Computer now drafts and revises saved skills from each user’s session history, capped at 10 per person. Perplexity’s own test shows people almost never write skills by hand.
- OpenAI lets Codex agents chase ChatGPT slowdowns, with engineers approving fixes. The Performance Factory logged more than 1,700 investigations in about three months. OpenAI’s own review found 30 percent of the root-cause calls were weak or unproven.
- 32 Claude agents beat one by 5x on NumPy, while GPT’s best soloist slipped. Replit says 32 Claude Opus 5.5 agents reached an 18.9x speedup against 3.7x for one. Its benchmark finds the gap comes down to where the agents talk: a shared board or direct messages.
Grading Their Own Homework: Benchmarks From the Vendors
Stripe, Google and Cloudflare each shipped something this time with numbers attached, or conspicuously without them.
- Stripe tested 14 coding agents on payments code. Most fail a strict bar.. Agents averaged 78.1 percent on partial credit but 29.4 percent when every criterion had to pass. Stripe’s own plugin lifted all 14 pairings, and Stripe is the one doing the grading.
- Gemini API lets agents load tool schemas only when needed. Google’s Tool Retrieval preview keeps big tool libraries out of the prompt until a task calls for them. Google has published no numbers on token savings or accuracy.
- Cloudflare opens weights on two small models that pull fields from documents. Clairic and Clairic-flash turn receipts, PDFs and tickets into structured fields. The preview endpoint now serves the larger model, which costs 2.4 times as much after December 1.
Locking the Cage: Testing Agents Without Trusting Them
One team built a fake internet to attack its browser agent safely. Another showed that sandboxed models can still whisper through shared hardware.
- Chrome built a fake internet so its browser agent can be attacked safely. Google’s Chrome security team says a sandboxed copy of the web, filled with generated sites and hidden attack payloads, lets engineers stress-test auto browse without touching real accounts.
- Sandboxed AI models passed signals through a shared inference server. In contrived LessWrong experiments, one model’s request size changed another’s arithmetic on a vLLM server, enough to win a color game and slip past a monitor. The authors call the setups contrived.
How Fast, How Far: Two Essays on the Road Ahead
One writer says AI already speeds up AI research tenfold. A DeepMind team says no single brake will stop what comes next.
- An essay argues AI already speeds up AI research tenfold. Linch says METR’s finding of a 2x capabilities jump understates the real research speedup and that governments should consider pausing frontier work. It is advocacy, built on redacted evidence.
- DeepMind researchers: superintelligence has four roads, so one brake fails. Tim Genewein and colleagues at Google DeepMind’s institute argue that compute caps alone cannot stop AI progress, and that labs and governments should start measuring the paths now.
Today’s Quick Hits
- Apple agent report hits Anthropic and Meta shares after hours. FX Leaders, citing a Bloomberg report, says Apple is hiring to build an agent that codes and drives browsers. Anthropic and Meta shares dipped after hours. Apple has announced nothing, and late 2027 is only a rumored window.
- Gemini on Mac can now click through your chores for you. Google’s Mac app gains an agent that works files, desktop software and websites from a hidden window. English and U.S. users only for now, and Google’s post gives no accuracy data.
- The creators of ETH are vetting the AI claim that it is wrong. UC San Diego and MIT theorists are coordinating outside checks of OpenAI and Anthropic claims that their models disproved the Exponential Time Hypothesis. Neither lab’s result is confirmed, and a community blog is open for findings.
- Apple researchers cut video generation to 36 ms a frame. Apple’s FastFrames replaces the video decoder and cache updates that ate over half the delay, putting generated video near the 50 ms window interactive use demands. The speed figure is on one Nvidia B200.