Parallel Fan-out / Gather

Definition

Split a task across multiple agents that execute concurrently, then an aggregator merges, votes on, or synthesizes the results.

Category: Information flow

When to use

Multi-perspective exploration, parallel retrieval, parallel code reading, log analysis, model ensembling.

When not to use

When the task is strongly order-dependent, when branches need shared mutable state, or when the budget is tight.

How to implement

1. Before fanning out, give each branch a distinct goal — avoid duplicated work.
2. Branches share read-only context by default, never the workspace.
3. The gather step does deduplication, conflict detection, and evidence merging — not naive concatenation.
4. Configure concurrency, timeouts, cancellation, and budget caps.

Minimal pseudocode

const branches = plan.branches.map(branch =>
  runWithTimeout(agentFor(branch).run(branch), branch.timeoutMs)
);
const results = await Promise.allSettled(branches);
return aggregator.run({ task, results });

Recommended trace events

• fanout.started
• branch.started
• branch.completed
• branch.timeout
• gather.started
• gather.completed

Common failure modes

• Multiple agents do the same work.
• The aggregator silently merges contradictory conclusions.
• Concurrent writes corrupt a shared workspace.

Implementation checklist

• Input/output schemas defined.
• Each agent's permission boundary defined.
• Every agent call carries a run id / trace id.
• Failure, timeout, cancel, and retry strategies defined.
• Context passed is the minimum required, not the full history.
• High-risk actions are gated by approval or a verifier.

References

• Google ADK patterns