agent-coordination-patterns

Choose an explicit coordination pattern first: sequential pipeline for dependent steps, parallel swarm for independent sub-tasks, or a dependency graph for mixed workflows. Assign each agent a non-overlapping scope and declare the merge strategy before spawning any agents.

Follow the conflict resolution ladder: re-scope with narrower boundaries, re-sequence so one agent critiques the other, apply a domain-owner authority rule, escalate to structured deliberation, or present both options to a human. Never silently pick one output without documenting why.

Use parallel agents when sub-tasks are independent and their outputs can be combined, such as auditing separate route groups. Use a sequential pipeline when each step genuinely depends on the previous step's output, like research feeding into planning and implementation.

Coordination fits agents with independent, complementary scopes working toward a shared goal, while deliberation fits fundamental disagreements requiring consensus before an irreversible decision. If coordination produces a deadlock, escalate to deliberation rather than force-merging.

Cap parallel agents at five per merge step. For six or more agents, use hierarchical fan-in where intermediate mergers consolidate groups before a final synthesizer combines results, preventing parent context from exploding.

Conflicting edits happen when agent scopes overlap. Re-partition work by file, domain, or layer so every partition has a single owner, define handoff contracts between agents, or switch to sequential execution when boundaries cannot be made clean.