Parallel Dispatch & Collection

Fan-out is the throughput multiplier. When you have N independent tasks, dispatching them in parallel finishes the work in the time of the slowest agent, not the sum of all agents. Five research agents investigating five competitors finish in 3 minutes, not 15. That time savings is the primary economic argument for multi-agent teams.

The dispatch pattern is straightforward. The coordinator identifies the independent work units, creates a prompt for each agent using a consistent template, and dispatches all agents simultaneously. Each agent works in isolation — no shared state, no inter-agent communication, no dependency on any other agent's progress. When all agents return, the coordinator collects results and begins synthesis.

Collection is where the coordinator earns its keep. Five agents return five results. Some are comprehensive. Some are thin. One might have failed entirely. The coordinator's job is not to concatenate — it is to evaluate, compare, identify gaps, and synthesize. "Agent 3's competitor profile is missing pricing data. Agents 1, 2, 4, and 5 all returned pricing. Re-dispatching Agent 3 with explicit pricing instructions." That is collection done right.

The synthesis step — combining N independent results into a coherent deliverable — is the highest-value work in the entire pipeline. The coordinator compares findings across agents, identifies patterns ("three of five competitors are moving to usage-based pricing"), surfaces contradictions ("Agent 2 reports Company X has 500 employees but Agent 4 reports 300"), and produces a unified analysis that no individual agent could have created. The synthesis is more than the sum of the parts.

1. 1. Identify Independent Work Units Map the full task and find the units that have no dependency on each other. If Agent B needs Agent A's output, they are not independent. If both agents can start with only the original brief, they are independent.
2. 2. Create Consistent Templates Each agent should receive the same prompt structure with a different target. Consistency makes comparison easier and ensures all agents return data in the same format.
3. 3. Dispatch Simultaneously Send all agents at once. Do not stagger. Do not wait for the first agent before dispatching the second. The time savings only materialize if all agents are running in parallel.
4. 4. Collect, Evaluate, Synthesize When results arrive, evaluate each for completeness and quality. Re-dispatch any agents that returned incomplete results. Then synthesize: compare, contrast, identify patterns, surface contradictions, and produce the unified deliverable.