Multi-agent systems are valuable when work needs distinct ownership, context, tools, policy, models, or independent parallelism. They also multiply prompts, calls, failure modes, approval surfaces, latency, and debugging cost. Start with one agent; adopt a pattern only when a measured limitation has a clear specialist contract and a simpler deterministic alternative is insufficient.
You will select a multi-agent pattern, identify coordination anti-patterns, and set an evidence threshold for adding another worker.
The harness—not a free-form agent—should own global policy, budgets, cancellation, shared state, and commit ordering.
Require a baseline failure and measurable hypothesis: smaller context should improve accuracy, separate tools should reduce unsafe exposure, parallel reads should reduce latency, or specialist model routing should reduce cost while meeting quality. Run an A/B evaluation with operational complexity included.
Failure injection: Let two agents edit the same file concurrently. Use the conflict to demonstrate why parallel analysis and serialized commit are safer.
Assert delegation choice, context boundaries, tool profiles, handoff depth, fan-out, deadlines, result schemas, conflict handling, and final ownership. Inspect traces for redundant work.
Compare single-agent, manager-specialist, and parallel-review designs on one task suite. Include coordination overhead, not only answer quality.