Mental Model

Multi-agentAgentA worker you delegate to: brief it, and it takes steps on its own. A chatbot answers; an agent acts.An LLM that runs tools in a loop toward a goal — it acts, checks the result, decides the next step, and repeats until done.Full definitionClose systems differ along four axes:

1. Decomposition authority — who creates and assigns work.
2. Communication topology — coordinator-only, peer-to-peer, or external bus.
3. Shared state — none, a task ledger, shared files, or workflow state.
4. Completion authority — who decides that the combined result is acceptable.

Product Topologies

Claude Code agent teams document independent sessions, a shared task list, mailboxes, and peer messaging, while remaining experimental. Codex documents parallel subagents whose results return through the primary coordinator. Copilot CLI Fleet documents orchestrated parallel work and sibling-agent coordination.

These are not equivalent implementations. A design that depends on peer negotiation cannot be moved unchanged to a coordinator-only surface.

Distributed-Systems Costs

Parallel agents introduce duplicated context, partial failure, stale state, conflicting writes, ordering problems, and reconciliation. The same design disciplines apply: idempotent tasks, explicit ownership, stable identifiers, timeouts, bounded retries, and a single source of completion truth.

The Crux

Use a team only when the task graph contains enough independent work to repay coordination overhead. If most steps depend on the immediately previous result, a single well-instructed agent is the better architecture.