Claude Code launches Projects; essay argues the agent harness is dissolving into weights
daniel_mac8 · x · 2026-09-18
Claude Code launches 'Projects': one interface where you describe an outcome and Claude autonomously spins up parallel cloud threads — an example of what the author calls the 'Attention Interface'.
His essay 'The Evolution of the Agent Harness' argues:
- Around Christmas 2025 agents suddenly started working — the confluence of better models and better harnesses; Transformer co-inventor Lukasz Kaiser also found the jump hard to pin down.
- From ChatGPT's 2022 'brain in a vat' (next-token prediction + RLHF), models keep absorbing the harness into their weights while engineers delete what got absorbed.
- What remains will be a harness for human attention rather than for the model.
More from AGI Musings
- Stanford's Chris Manning: universities fund CS like it's 2005 and are losing AI professors to frontier labs — stanfordnlp · 2026-09-18
- Debate: Is AI Eval/Safety Work Really Only for a Tiny Talent Pool? — JacquesThibs · 2026-09-18
- Viral greentext skewers effective altruism: 10^50 unborn people as the perfect constituency — banteg · 2026-09-18
- New LessWrong Essay 'The Obliqueness Thesis' Reopens The Orthogonality Debate In AI Alignment — jessi_cata · 2026-09-18
- LBC Call-In Segment Shows the Public Thinks About AI Smarter Than Many Experts — ShakeelHashim · 2026-09-18
- Educators Debate Banning AI Writing vs. Testing Authors' Actual Understanding — Dr_Atoosa · 2026-09-18