AI Teams

A coordination system

A guide to coordinating multiple Claude agents on a single problem — covering the routing logic, parallel dispatch patterns, and supervision structures that let agents work together without stepping on each other.

If you have hit the limits of what a single session can do, this gives you the architecture for going further.

Five areas, one coordinated system

• agent-routing — Assigning work to the right agent: model-selection signals (Haiku for lookups, Sonnet for standard tasks, Opus for architecture and deep analysis), task-complexity scoring, context-size limits per tier. The routing schema that makes the assignment programmable rather than manual.
• parallel-dispatch — Running multiple agents concurrently: spawning agents with independent prompts, collecting results with a barrier, merging into a unified output schema. Collision patterns to avoid when agents write to shared state. The fan-out / fan-in loop with deduplication before the merge pass.
• shared-state — How agents pass context across steps: the findings_handoff JSONL schema (agent_id, task_id, findings[], status, timestamp), append-only writes so no agent overwrites another’s output, and the reader pattern that loads all prior findings before generating new ones.
• supervision-loops — Catching agent failures and rerouting: exit-code checks after each call, retry logic with exponential backoff (1s, 2s, 4s, max 3 retries), failure escalation to a supervisor agent that diagnoses and re-scopes the task. The loop that never blocks on a single stuck agent.
• trust-architecture — Permission tiers by capability level: read-only agents (no file writes), restricted agents (write to designated output dirs only), trusted agents (full tool access). Enforcement at the hook layer — so a misconfigured agent can’t exceed its permission tier regardless of what its prompt says.
• parallel-dispatch — Running multiple agents concurrently and merging their output without collisions.
• shared-state — How agents pass context to each other across steps without duplicating work or losing findings.
• supervision-loops — Catching errors and rerouting failed steps without manual intervention.
• trust-architecture — Which agents get which permissions, and how to keep high-capability agents from running unsupervised.

A preview

The bottleneck in most agent systems isn’t the model — it’s the plumbing. Routing logic, state handoff, and supervision are where teams break down, and they’re all human design problems.

Parallel dispatch is only useful if the merge step is designed before the agents run. Most coordination failures happen at the join, not the fan-out.

Users ready for multi-agent work

Claude users who have outgrown single-session workflows and are ready to coordinate multiple agents on complex, multi-step jobs. Prerequisites: comfortable with Productivity and Automations.

A longer excerpt is available to newsletter subscribers.

Buy on Amazon