Problem
AI coding assistants are great in a single editor tab and much less great across a real product: several repositories, shared conventions, architecture decisions that span services, and a backlog that never sleeps. Most teams stop at “autocomplete with extra steps”.
What I built
A dedicated AI agent server running persistent sessions, organised like an engineering team:
- Repo owners. Long-lived Claude Code sessions, each responsible for one or more repositories. They hold the context of that codebase over time.
- Coordinators. One or two central sessions that plan work, split it across repo owners, and keep architecture and conventions consistent between them.
- The whole loop. Agents handle tasks, documentation, architecture notes, implementation, tests and convention checks, and hand results back for human review.
Decisions that mattered
- Persistence over prompts. Long-running sessions per repo beat starting from zero every time. Context is the scarce resource.
- Clear ownership. One agent, one area. The same rule that makes human teams work stops agents from stepping on each other.
- Humans approve, agents execute. Review gates stay with people; the agents do the grinding.
Results
The same few people cover far more ground: documentation stays in sync with code, conventions are applied consistently, and routine implementation and testing run in parallel instead of queueing behind one engineer.
Lessons
The model is rarely the bottleneck. Structure is: ownership, coordination, guardrails and good review habits. That’s exactly what I help clients set up.