Anthropic Engineer Proposes 'Loop Engineering'

blaizedsouza · x · 2026-07-18

A senior Anthropic engineer shares the concept of 'Loop Engineering' for building agentic systems, advocating for building systems that prompt themselves instead of manual prompting. The system operates on a **Plan → Discover → Build → Verify → Repeat** loop, with five key steps per cycle: - **Discover**: Autonomously finding work (like failed CI, unresolved issues) rather than passively receiving tasks. - **Handoff**: Assigning a dedicated git worktree to each task to avoid parallel agent conflicts. - **Verify**: Introducing a second agent to act as an 'adversary', reviewing the code with the assumption that it has flaws. - **Persistence**: Writing results to disk to prevent loss during context refreshes. - **Scheduling**: Using cron jobs to wake up the agent and close the loop. The core insight is that agents often 'grade their own homework' during self-evaluation, making external verification mechanisms essential.

Related event: Anthropic Engineer Introduces 'Loop Engineering' for AI Agents(4 posts)→

Original post →

More from coding & agent

coding & agent channel →