Person/Dan Shipper/Blog/25/12/11/Compound Engineering - How Every Codes With Agents

• Compound Engineering: How Every Codes With Agents

  • Co-authored with Kieran Klaassen (every.to/@kieran_1355); updated 2026-04-06.

  • Summary

    • At Every, essentially 100% of code is agent-written; engineering shifts from typing to orchestrating agents in a plan → work → review → compound loop.
    • Compound engineering expects each feature to make the next feature easier by persisting lessons (bugs, failed tests, review insights) for future agents—while codebase complexity still grows, agent knowledge of the codebase grows faster.
    • Five in-house products run primarily by single owners serving thousands of users; claim: one developer with the right system can match ~five developers of a few years ago.
    • Open-sourced workflow: Every/GitHub/compound-engineering-plugin.

  • Notes

    • Plan: agents study repo structure, commit history, and external best practices before planning—most human time lives here.
    • Review: engineer judges output and meta-lessons, not just line-by-line nits.
    • Compound: capture what was learned so the next agent session inherits it (feeds into CLAUDE.md-style context and team memory).
    • Tool-agnostic framing: Claude/Code primary internally; also Factory Droid and Codex/CLI.

  • Links

    • Engineer/ing/Compound
    • Every/Blog/Source Code/25/08/18/My AI Had Already Fixed the Code Before I Saw It — earlier narrative on the same idea.