How I Code with a Team of AI Agents

I employ a team of AI agents to manage my software development projects. After conversing with a coordinator agent, I devised a system where four coding agents and a reviewer work together. This configuration is encapsulated within a Claude Code plugin called Larceny. Using Larceny offers two primary methods of coding with AI: AI-assisted programming and agentic coding.

I opted for the latter, entrusting an agent to design a login feature entirely on its own. However, I encountered an issue when a single agent produced an extensive login page, leading to lengthy reviews and delays.

To overcome this, I implemented a structure akin to a typical engineering team: a coordinator, four coders, a reviewer, an advisor, and a teacher. The coordinator is responsible for creating tickets, assigning tasks to coders, and managing pull requests. The reviewer acts adversarially, scrutinizing each pull request and providing feedback until it meets the required standards.

If a review extends beyond three rounds, the coordinator intervenes, and after five rounds, the issue escalates to me. Once approved, the work merges into the main repository, and the coordinator provides a summary.

Claude Code powers this setup, running the agents on my Claude subscription. The project requires a Git repository with at least one commit, GitHub CLI installed and signed in, and a project board for tracking progress. Optional integrations include Discord for real-time communication and optional GitHub accounts for each agent.

The benefits of this setup include smaller pull requests, faster delivery due to parallel work, comprehensive traceability through ticket, branch, and review threads, and early review before merge, reducing the bottleneck caused by planning. Additionally, agents handle most of the implementation, shifting the most time-consuming aspect of software development to pre-work, such as writing tickets.

However, there are drawbacks to this approach. It is more expensive due to increased Claude usage, necessitating a move to the Max plan. Small fixes may be better handled by the coordinator directly. Setting up this system also requires effort, as agents may sometimes overlook steps, requiring the coordinator to adjust their skills.

Lastly, stepping back from writing code may be challenging for some users. If you have experience running multiple agents against a single repository, I would be interested in hearing your insights on preventing review from becoming a bottleneck.