vault / agents / multi-agent.md

Multi-agent systems

Agents & LLMs 90 words 2 outgoing 1 backlinks

Before adding agents, ask whether the work splits into independent pieces that each need a large separate context and can run in parallel. Then a team can pay off: Anthropic’s research system runs parallel subagents, beats one agent on broad queries, and costs roughly 15x the tokens. Dependent steps - clarify, retrieve, answer - are a pipeline, not a team: one agent, explicit stages, plain code between them. Chained agents in the live path are where the latency and maintenance pain shows up.

Source: Anthropic on their multi-agent research system.

LINKS IN THIS NOTE

[[ Evaluating an agent ]] A fluent answer proves nothing, so evaluate the loop rather than the vibe. Compare relevance, retrieval recall, groundedness and latency against a baseline, and read examples next to the scores: grounded, useful and found-the-right-document are three different failures. Offline golden... [[ Observability ]] Instrument so a stranger can reconstruct any request after the fact: the question, what was retrieved, the answer, and the configuration in force. Dashboards show a pattern; a concrete replay tells you which component or document to fix. Gotchas: logs...