ReAct to Production Agents: The 7 Layers Between Model and Result
blaizedsouza · x · 2026-11-03
A thread recommends two core resources for understanding and building AI agents:
- ReAct (Princeton & Google) — called the most influential agent paper online. Its think→act→observe→think-again loop powers Claude Code, Codex, and Cursor; reasoning alone hallucinates, acting alone gets lost, and together they fix each other. With just one or two examples it beat trained agents by 34% on ALFWorld and 10% on WebShop. The thread maps key pages: Figure 1, the loop mechanics, ReAct vs CoT tables, and a human fixing an agent by editing one thought.
- "How to Build a Production AI Agent" — most agent projects start with the strongest model, a long prompt, and a demo, then break within a week (numbers not matching sources, lost context); the guide lays out the 7 layers between the model and the result.
More from coding & agent
- OpenAI's custom source-control system Pando moves nearly a petabyte of Git data daily, cutting push latency 5x — Aizkmusic · 2026-11-03
- Karpathy: one prompt to Claude Code deploys a YouTube thumbnail-swap browser extension — amaarora · 2026-11-03
- Cherry Pi proposes layered harness architecture: agent loops compose with governed workspaces — sull · 2026-11-03
- $200 Claude Max runs overnight multi-subagent Opus 5.5 job, burns only 6% of weekly quota — lxfater · 2026-11-03
- Agent claimed it fixed the security bug — the original PoC still fired: verify the trigger, not the diff — AIGODSEND · 2026-11-03
- Google & HKU's VeriCUA cuts false task-completion rate to 5.4% by checking saved files, not agent claims — rohanpaul_ai · 2026-11-03