AI-NATIVE ENGINEERING
2025 · multi-agent orchestration
nine agents, four codebases,
one product.
A university faculty app with a backend, a dashboard and a mobile app, shipped in two weeks. 10,000+ people use it every day.
built before codex and claude shipped similar capabilities
problem
Give 9 AI agents access to everything and they overwrite each other's work. With no single owner per codebase, every agent you add means more conflicting edits.
architecture
Each codebase has one agent with write access. The others can only read. All coordination goes through shared docs. API contracts and specs keep the agents in sync, so none of them touches another's code.
outcome
No cross-repo conflicts in the two-week build. Each agent works on its own inside its codebase. I review what it produces instead of watching every session.
9agents orchestrated
4codebases coordinated
2wksto ship
10k+daily users
0cross-repo conflicts
decision log
- 01One agent with write access per codebaseAgents with access to everything overwrite each other's work.
- 02Shared docs as the only way agents coordinateAPI contracts and specs keep them in sync without any agent touching another's code.
- 03QA, security and performance agents can only readThey can check every codebase without being able to change one.
- 04Risky changes come to meI review results and escalations instead of watching every agent session.