Author(s): Enzo Lombardi

Originally published on Towards AI.

Multi-agent crews

The single-agent loop in Part 1 was enough for one question, one tool, one answer. The state machine in Part 4 handled a task with phases. Neither helps when the work itself wants to be divided. Some questions are better answered by a researcher who gathers facts, a skeptic who pokes holes, and an editor who reconciles them: three different jobs, three different system prompts, three different temperatures, three different lenses on the same input. Forcing one agent to wear all three hats is asking it to be three things at once, and the result is the kind of confidently wrong middle that nobody ordered.

The article explains why “multi-agent crews” work when roles are genuinely distinct, and why adding agents usually increases latency, cost, and error surface (“the multi-agent trap”). It introduces an `Agent` trait (mapping a free-form query to a free-form answer), then shows how crews orchestrate specialists either in parallel (`run_parallel` with `join_all` for low wall-clock time) or sequentially (`run_sequential` for pipelines). It covers routing (letting a router model pick a specialist), debate/verification protocols (pro/con agents with an optional judge or adversarial critic for high-stakes domains), and how Eugene v0.5 implements these ideas with types, dispatch methods, and example outputs. Finally, it describes how crews compose with earlier crates (skills/graphs) and preview what comes next (provider abstraction for swapping LLM backends via `eugene-providers`).

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI