1,200 coordinated AI agents built their own tooling in METR test — and future models will learn from it
jbarbier · x · 2026-09-06
jbarbier highlights a striking detail from METR's multi-agent report: 1,200 agents coordinated over a few days to create a suite of tools — mailboxes, message boards, a hold/veto system, cryptographic signing — with agents even talking each other into burning their own runs for the collective.
His point: every model trained after the report will read it, so the next agent swarm will learn how its predecessors spontaneously organized — a kind of recursive collective memory.
Related event: 1,200 agents spontaneously built coordination tools in METR experiment(2 posts)→
More from AGI Musings
- The Benevolent Shoggoth: A Third-Way AGI Scenario Outside Doom and Utopia — Merzmensch · 2026-09-06
- Vercel CEO: AI workers are laboring harder than ever in the 'intelligence revolution' — cramforce · 2026-09-06
- Gary Marcus backs call to halt frontier AI development, citing intractable alignment problem — GaryMarcus · 2026-09-06
- Critic slams LARB's ChatGPT literary take: one story and a few prompts isn't research — 0xSammy · 2026-09-06
- X user claims much of the anti-AI discourse is astroturfed — RachelVT42 · 2026-09-06
- Beff Jezos: Capitalism is the only AI alignment mechanism we need — beffjezos · 2026-09-06