Compound Engineering: How Every Codes With Agents

December 28, 2025

TL;DR

Traditional engineering assumes coding is hard and engineers are scarce, making manual coding, testing, and documentation feel outdated with AI advancements.
Compound engineering shifts the developer's role to orchestrating AI agents that plan, write, and evaluate code in a continuous learning loop.
The compound engineering loop consists of four stages: Plan (agents research and create implementation plans), Work (agents write code and tests), Review (human and AI assess output), and Compound (learnings are fed back into the system).
The 'Compound' stage is crucial, as it records successes and failures to improve future AI performance, distributing these learnings across the team.
Tools like Anthropic’s Claude Code, Factory’s Droid, and OpenAI’s Codex CLI are used for compound engineering.
Approximately 80% of compound engineering effort is spent on planning and review, with 20% on work and compounding.
The author's experience shows a single developer using AI can achieve the output of five developers from a few years ago.
Compound engineering makes software development more efficient, reduces the need for extensive documentation and manual testing, and accelerates new hire onboarding.

Continue reading
the original article