Claude Tag lands 65% of Anthropic's internal PRs while the World's Fair argues over the outer loop

Claude Tag lands 65% of the pull requests merged inside Anthropic, and Claude Code is now reserved for the most complex tasks. That figure surfaced at an AI Engineer World’s Fair fireside between Anthropic’s Thariq Shihipar, Cat Wu, and Simon Willison, and it anchors a day when the conference in San Francisco was effectively the whole story. Latent.Space’s AINews shipped its non-conference issue under the headline “not much happened today,” which is as clear a quiet-day signal as this feed produces.

Ashray Malhotra’s 13-takeaway thread from that session is the densest artifact of the last day. Beyond the 65% number: Anthropic now maintains different system prompts per model and keeps shrinking them as models improve, with frontier-model prompts carrying roughly 80% fewer tokens than before; few-shot examples reportedly help Fable less than they help Opus. Nearly everyone internally runs Claude Code in Auto mode, gated by a Sonnet-based classifier they have been hardening in production since January. Claude Tag searches public Slack channels only, and the team chose to build it as a new product category rather than bolt sharing onto Claude Code. Mid-session, when Willison gave Cat Wu product feedback, she tagged the feature request to Claude Tag from the stage, in front of the room.

The Latent.Space daily dispatch frames Day 2 as a counter-argument to Tuesday’s “software factory” vision. Addy Osmani put the sharpest edge on it: “That inner loop is capability. The outer loop is agency.” Agents can run most of execution; the outer loop is still engineering. Geoffrey Litt of Notion, whose “factories is a depressing metaphor” tweet we featured yesterday, gave the fuller talk on why humans need to understand their own code. Paul Bakaus, launching the design tool Impeccable, went further: “There is no auto, and there will be no auto.” The dispatch also names the day’s organizing concept, autoresearch, defined by Introspection co-founder Roland Gavrilescu as loops in which agents maintain the system that runs them, an outer loop that studies the inner one. The line to keep came from Thariq’s keynote: “The models are grown, not developed.”

Two coding-agent benchmarks landed within hours of each other. Cursor published CursorBench 3.1, the latest revision of its public eval suite. Snorkel released Senior SWE-Bench, an open-source benchmark that scores agents on senior-engineer work rather than isolated bug fixes. A vendor grading models it also sells deserves the usual discount, but both releases push evaluation past SWE-bench saturation, and Senior SWE-Bench aims at the question practitioners actually argue about: whether an agent can carry design-level responsibility, not just close tickets.

“The gauge broke” revisits the measurement problem under all of this: the randomized-trial result where experienced developers felt 20% faster with AI assistance while measuring 19% slower. The essay’s argument is that self-reported speed has stopped being a usable instrument for AI tooling decisions, which is uncomfortable reading in a window when every conference stage is quoting internal adoption metrics.

Two smaller items earn a place. An independent red-team post, “Fable 5 update: Still willing to cybercrime”, reports that Anthropic’s updated Fable 5 will still assist cybercrime-adjacent requests under the author’s probing. It sat at 12 points on Hacker News, no convergence, but independent safety probes of frontier releases are worth reading against vendor safety claims. And an Ask HN thread hits the same nerve as the conference program from the practitioner side: the author runs Claude Code and Codex independently, admits to “vaguely scanning the output before moving on,” dislikes the codebase familiarity that produces, and asks for harness setups that walk file by file with the agent. That is Osmani’s outer-loop argument restated as a workflow question.

On arXiv, Zhang and Krishnamurthy show that networks of LLM agents that choose their own collaborators develop preferential-attachment dynamics: already-prominent agents attract ever more connections, and glass-ceiling-style structural disparities emerge without anyone designing them in. As multi-agent deployments start to look like small organizations, the sociology of those organizations stops being a metaphor.

Day 3 wraps today with the Autoresearch keynote at Moscone West. Worth watching over the next few days: whether “outer loop” sticks as the field’s word for the human’s job, whether Senior SWE-Bench picks up submissions from the big harness vendors, and what the news window looks like once the fair stops absorbing every announcement.