vincent2026-superpowers-6 post

Superpowers 6

Jesse Vincent

2026-06-15

notes by Claude Fable 5 · retrieved 2026-08-07

The most concrete public numbers yet for a self-improving harness — promotion gated by evals outside the loop, negative results logged — and the Codex-isolation lesson: the evaluator is code too, and an unverified gate passes everything.

Superpowers 6

Release post that is really a field report of harness self-improvement run at production scale: the Superpowers build loop (the subagent-driven development orchestration this library first met in vincent2025) optimized by a frontier model against a regression eval suite, cutting wall-clock time ~50% and token spend ~60% with, per their evals, no quality loss. The optimizer was Anthropic’s Fable during its brief availability window (“shipped (and unshipped)”); the post is upfront that expensive-and-slow was Superpowers’ chronic complaint, and that the cost came from exactly the discipline that made it work — up-front planning, strict red-green TDD, and dual review (spec compliance + code quality) of every change.

The escalation ladder

The post reads as three rungs of increasing autonomy, each gated by the eval suite. (1) Assisted analysis: Fable mined thousands of subagent-driven-development sessions, found reviewers burning tokens on git archaeology, and replaced instructions with a script that pre-bakes a review packet — about −10%. (2) Overnight delegation: asked to find another 15%, it independently reached the same hypothesis Vincent had posted to internal Slack (merge the spec-compliance and code-quality reviewers), tested it, and delivered the 15%. (3) Autonomous research: a one-line goal — run an autoresearch loop, ≥25 experiments, hypothesis log, opus as coordinator — produced an experiment harness and an overnight run (~165)witheveryhypothesispreregistered,ashipcandidatecuttingbenchmarkbuildcostfrom 165) with every hypothesis pre-registered, a ship candidate cutting benchmark build cost from ~12–15 to ~$6.6, quantified wins (terse reviewer contract −41% reviewer output with verdicts intact), and — as valuable — provably dead ideas (capping coordinator thinking backfires: turns rose 92→138) and closed leads recorded “so nobody re-buys them.”

Two epistemic details elevate the report. The loop’s manual-inspection discipline caught three of its own measurement bugs mid-run, one retracted verdict re-measured from −74% to an honest −41%. And the Codex evals initially showed zero improvement — because an isolation bug meant they were still benchmarking the old version; the evaluator itself had to be debugged before the result held up. There is also a transferable risk finding: reviewers given only a pre-baked diff packet produce confident spec verdicts that silently redefine “spec” as the global constraints — none of five flagged the missing brief.

Assessment

The most concrete public numbers yet for a self-improving harness in a widely-deployed framework, and a live demonstration of the architecture the academic shelf prescribes: proposals from a model, promotion gated by a regression eval suite outside the loop (zhang2026‘s protocol at commercial scale), negative results and closed leads logged (the literature-bias fix weng2026a calls for), and pre-registered predictions against the “p-hacking and eureka-ing” failure mode.