Addy Osmani's Data: Three-Agent Teams Beat Solo Agents. And Your LLM-Generated AGENTS.md Is Hurting You.

Addy Osmani's O'Reilly CodeCon talk, published March 26, gives us the first real empirical data on multi-agent coding that I trust. Not vendor benchmarks. Not cherry-picked demos. Controlled measurements across real development tasks.

The headline finding: three focused agents consistently outperform one generalist agent working three times as long. Not sometimes. Consistently. The sweet spot is 3-5 agents. Token costs scale linearly with agent count, but completion quality scales super-linearly. Past 5 agents, you hit coordination overhead that eats the gains.

Three coordination patterns, ranked by sophistication. First: subagents, where a lead agent delegates parallel subtasks. Second: Agent Teams, with a shared task list, peer messaging, and file locking. Third: hierarchical subagents, where feature leads spawn their own specialists. Each pattern fits different problem shapes, and Osmani gives specific guidance on when to use which.

Here's the finding that stopped me. LLM-generated AGENTS.md files offer no benefit. They marginally reduce success rates by about 3%. Developer-written context files provide roughly 4% improvement. That's a 7-point swing between "had Claude write my context file" and "wrote it myself." The LLM doesn't know what matters about your project. You do. The context file is where your taste shows up. Outsourcing it to the same tool you're trying to guide is circular.

This connects directly to the reads-per-edit ratio that AMD's Stella Laurenzo documented. Healthy agents read 6.6 files per edit. Degraded agents read 2.0. Osmani's multi-agent teams naturally produce higher reads-per-edit because each agent focuses on understanding its piece deeply before changing anything. The single generalist tries to hold the whole codebase in context and edits with insufficient understanding.

For builders: if you're using a single Claude Code session for everything, try splitting into three focused sessions. One for the data layer, one for business logic, one for UI. Give each a dedicated CLAUDE.md scoped to its domain. Write those files yourself. Track your reads-per-edit ratio in your session JSONL files as a quality signal.