CORE GUIDE
Coordination overhead
Multi-agent systems pay real costs for handoffs, duplicated context, conflict resolution and verification; parallelism is valuable only when independent work saves more than coordination consumes.
Mental model
Coordination overhead is the extra work created by decomposition itself: assigning tasks, transferring state, reconciling outputs, handling partial failure and proving that the combined result is coherent.
Why it matters
Adding agents can improve specialization or throughput, but each boundary creates latency, token use and new failure modes. If subtasks are strongly dependent, agents spend time explaining state to one another or repeating work. A multi-agent design should therefore beat a simpler loop on measured outcomes rather than on architectural appearance.
01
Account for the cost of every boundary
Before splitting work, identify independent subproblems and the data each requires. Estimate duplicated context, serialization, queueing, merge and verification cost. Give handoffs explicit schemas and ownership. Compare the design with a single-agent or sequential baseline under the same task set, measuring quality, wall-clock time, compute and recovery complexity.
02
Example: three agents write one tightly coupled migration
A planner, database agent and API agent work in parallel on a schema migration, but each needs the others' latest interface decisions. They repeatedly exchange updates and produce conflicting assumptions. A single sequenced workflow finishes faster. Parallel agents become useful only after the migration is decomposed into independent modules with a stable contract.
Common failure modes
- Splitting work by role names instead of actual dependency boundaries.
- Measuring parallel worker speed while ignoring merge and verification time.
- Allowing shared mutable state with no ownership or conflict policy.
Engineering heuristics
- Use a simpler baseline in every orchestration evaluation.
- Parallelize only work with genuine independence or latency hiding.
- Include coordination and recovery cost in the success metric.
Takeaways
- 01Decomposition creates overhead as well as opportunity.
- 02Parallelism must earn its coordination cost.
- 03Topology should be justified by dependencies and measured outcomes.
Reading evidence
This records actions you actually took; it does not claim mastery, proficiency, or certification.
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.