CORE GUIDE
Parallel agent work
Parallelism helps when subproblems are genuinely independent or provide independent verification; otherwise coordination cost can exceed the speedup.
Mental model
Parallel agents are concurrent workers over explicitly partitioned state. They earn their complexity when work can proceed independently and their outputs can be reconciled through a clear contract.
Why it matters
Launching several agents is easy; obtaining trustworthy speedup is not. Agents can duplicate work, edit the same state, propagate inconsistent assumptions or require expensive synthesis. Parallelism should therefore be introduced after decomposition, ownership and merge rules are explicit. Independent verification is often a stronger reason to parallelize than simply creating more personas.
01
Partition work and define the join
Identify tasks whose inputs do not depend on one another's intermediate reasoning. Assign each worker a bounded artifact or decision, give it only the relevant context, and define how results will be joined. Shared mutable state should have ownership or serialization rules. The coordinator then verifies completeness, conflicts and evidence quality before accepting the combined result.
02
Example: parallel repository investigation
One worker traces an API failure through backend code while another independently inspects frontend request construction. Both return file references, hypotheses and evidence rather than editing the same files immediately. A coordinator compares the findings and assigns one implementation path only after the shared failure boundary is understood.
Common failure modes
- Parallelizing tasks that depend on one another's unresolved intermediate state.
- Letting multiple agents write the same files or records without an ownership rule.
- Measuring only wall-clock speed while ignoring duplicated tokens, conflicts and synthesis cost.
Engineering heuristics
- Parallelize independent evidence gathering before parallelizing shared-state mutation.
- Give each worker an explicit output contract and ownership boundary.
- Measure total coordination cost and defect rate, not just elapsed time.
Takeaways
- 01Concurrency is useful only when decomposition is real.
- 02The join step is part of the architecture, not an afterthought.
- 03Independent parallel checks can improve reliability even when they do not reduce latency.
Used in
This Concept is reused across these canonical learning paths.
Related concepts from the Knowledge Graph
These relationships come from the canonical graph, not a separate Guide taxonomy.