liveBook · Manning

Fan-out/gather

As explained in Designing AI Agents

A coordination pattern that distributes work across multiple agents or workers and then aggregates their results.

Fan-out/gather distributes independent work to parallel workers, then reconnects their results at an accountable gather point. It is useful when branches can progress independently and their outputs can be combined without repeating the work. section 8.3

Parallel workers add search capacity, but also introduce coordination costs: lost constraints, duplicate work, incompatible answers, and integration effort. Compare the team’s task outcome with a credible single-agent baseline before assuming that more workers help. section 8.0 section 8.10

Define the split and the gather before dispatching workers:

  • Partition by independent units such as companies, file sets, jurisdictions, data shards, or distinct evidence searches. Different labels alone do not prevent duplicated reasoning. section 8.3
  • Set worker tool and write boundaries, a concurrency limit, a shared return schema, and a policy for failed workers. section 8.3
  • Assign an owner for the merged result. Define how to deduplicate records, check coverage, surface contradictions, and preserve consequential minority findings. section 8.3 section 8.3.2
  • Measure missing branches, duplicate work, contradictions, merge effort, latency, and total tokens alongside the task outcome. section 8.3 section 8.10

Board-member research

A lead agent divides a company set among workers, each responsible for finding board members and citations for assigned companies. A return record can include the companies, names, primary-source citations, dates checked, and unresolved conflicts. The gather step deduplicates names and checks coverage without discarding conflicting evidence. section 8.3

In parallel code review, workers can return claims, file locations, evidence, severity, proposed actions, and uncertainty. Gather findings by underlying cause, display incompatible recommendations, and retain a material finding even if only one worker discovered it. section 8.3.2

Table 8.4 Choose the gather rule by result shape
Result shape
Gather rule
What must remain visible
Additive evidence
Concatenate under one schema
Source and branch provenance
Comparable alternatives
Score or vote under a stated rule
Minority result and scoring basis
Shared structured records
Merge and deduplicate by cause
Conflicts, missing fields, and uncertainty
Consequential contested claim
Route to Adversarial Review
Claim, evidence, judge, and stop rule

Choose a gather rule for the result shape while retaining provenance, uncertainty, or minority evidence. section 8.3.2

How Designing AI Agents covers it

How deep it goes
Specifies partitioning, worker boundaries, return schemas, failure policy, merge rules, and operational metrics.
Examples
Uses board-member research and parallel code review to illustrate structured results, coverage checks, and disagreement preservation.
What you can do after
Apply the pattern by partitioning independent work, defining worker and merge policies, and measuring results against a single-agent baseline.
How it connects
Connects parallel coordination to MapReduce, fork/join, scatter-gather, and adversarial review.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.