Design multi-agent systems and orchestration strategies
Use multi-agent orchestration only when work parallelises or exceeds one context window, with a coordinator that owns decomposition, synthesis and side effects, scoped briefs for subagents, and effort scaled to query complexity.
Key points
- 1
Multi-agent systems excel at tasks with heavy parallelization, information that exceeds a single context window, and many complex tools. They are not a way to make a small sequential pipeline more robust.
- 2
Anthropic measured multi-agent research runs at roughly 15x the tokens of a chat; the economics only work when the task's value justifies that spend.
- 3
A coordinator (lead agent, orchestrator) plans the decomposition, gives each subagent a distinct objective, scope, tool guidance and output format, and owns the final synthesis. Peer agents messaging freely produce loops and ownerless answers.
- 4
Subagents run in their own context windows and return condensed findings, not raw pages; this is what keeps the coordinator's context from overflowing and lets the system scale past one window.
- 5
Vague delegation causes duplicated and off-target work; the fix is a better brief, not more workers.
- 6
Scale effort to query complexity: simple fact-finding needs one agent with a few tool calls, direct comparisons a few subagents, and only complex research more than ten subagents with clearly divided responsibilities. Put these heuristics in the lead agent's prompt to stop over-spawning.
- 7
Side effects belong with the coordinator (or a single writer): workers returning structured results that the coordinator deduplicates, merges and writes as validated transactions avoids conflicting writes and half-written records. Prompt instructions to "check first" are racy.
- 8
Long-running multi-agent runs need durable state: persist subagent results outside the context and checkpoint so an intermittent failure retries from where it stopped rather than restarting the whole run.
- 9
Citations are best added as a dedicated final pass that maps each claim in the synthesised report to its collected source.
- 10
Signals against multi-agent: cost minimisation, tight latency for simple lookups, tightly coupled edits to shared state in strict order, and stylistic reasons such as personas.
- 11
When declining multi-agent, the reason is the task's shape (sequential, small context, no parallelism), not its industry.
Read the source
Test yourself on Design multi-agent systems and orchestration strategies
Ten questions, with the answer and explanation after each one.