Agentic Research

Orchestrator–worker

A lead decomposes the task, dispatches workers, and collects results.

Shape:Hub and spoke; the lead holds the whole picture

When to use it

Long tasks needing several skills, where the decomposition itself requires judgement. The lead decides what to do; a worker finishes one thing.

When not to

When the task fits in one pass. Adding an orchestration layer only adds coordination cost and failure surface.

How it fails

  • The lead's context balloons. Every worker report lands in the lead, so the lead hits the context ceiling first.
  • Workers misreport. The lead can only trust the report, and the report is written by the worker — omission in manual reporting is the norm, not the exception.
  • Unclear task boundaries let two workers edit the same place, and the later write silently overwrites the earlier one.

Design notes

Worker write scopes must be disjoint and recorded. The lead has to be able to verify a worker's report, not just relay it. Split parallel work by file or resource, never by topic — splitting by topic always collides.

Origin

2025-06-13

Anthropic publishes the engineering internals of its multi-agent research system

https://www.anthropic.com/engineering/built-multi-agent-research-system

Related articles on this site