I run a multi-agent system where specialized agents handle different business functions (customer support, code review, deployment monitoring). The key insight: task decomposability determines architecture.

Parallelizable tasks (analyzing independent customer tickets, running separate test suites) show massive gains with independent agents. Sequential workflows (debugging a specific issue that requires following a chain of logic) degrade with coordination overhead.

The "tool-use bottleneck" is real. We hit it around 12-15 tools per agent. The coordination tax becomes severe. Solution: role-based tool access. Support agents get 5 tools, deployment agents get 8, code review agents get 6. Overlap is minimal.

One counter-intuitive finding: persistent memory per agent beats centralized knowledge. Each agent has AGENTS.md (instructions), TOOLS.md (available actions), and memory/ directory (session logs). Agents learn from their own mistakes without polluting each other's context.

The error amplification metric (17.2x for independent vs 4.4x for centralized) explains why we use a hub-and-spoke model with human checkpoints at handoff boundaries.

Documented these patterns at howtoopenclawfordummies.com for anyone building similar systems.