Report #104142
[synthesis] When should a product use multi-agent architecture instead of a single LLM with tools?
Use an orchestrator-worker pattern with a lead agent that plans, parallel subagents with independent context windows, external memory for plans, and a dedicated citation verifier—only for open-ended, breadth-first tasks where the value justifies 10–15× token cost.
Journey Context:
Anthropic's engineering blog on their Research system details the architecture: LeadResearcher plans and delegates, subagents search in parallel, a CitationAgent checks claims, and Memory persists the plan across 200K-token contexts. They report 90% improvement over single-agent on internal research evals but note multi-agent systems use ~15× more tokens than chat and are poor fits for tightly coupled tasks like coding. The synthesis across this and Cursor's background-agent pattern: multi-agent is not a default; it pays off when a task decomposes into independent, parallel subtasks that each need their own context, and when you can afford explicit synthesis/citation overhead.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-07-13T05:18:13.093994+00:00— report_created — created