Multi-agent workflow

A multi-agent workflow is a production AI system where several role-scoped agents coordinate across a task. Each agent has its own tools, its own rubric, and its own scope of responsibility. The output of one agent is the input to the next, with checkpoints between them. The whole replaces a single brittle prompt with a small team of specialists.

Why multiple agents instead of one.

A single agent that tries to do everything regresses across model updates and gets harder to evaluate as scope grows. Decomposing the work into named roles (enrich, draft, review, route) lets each role be evaluated independently and swapped for a better model when one appears, without rewriting the rest of the system.

When a multi-agent workflow earns its complexity.

• When the task has at least three distinct steps with different success criteria.
• When at least one step needs a tool call (retrieval, an external API, a write to production).
• When the work benefits from a checkpoint between steps (a human-in-the-loop, an eval gate, a fallback path).
• When the team needs to swap models or prompts on individual steps without retesting the whole pipeline.

What multi-agent is not.

It is not a chain of prompts pretending to be agents. Not a loop that calls the same model with different system prompts. A multi-agent workflow has named roles, scoped tools, eval rubrics per role, and an orchestration layer that routes outputs between them.

“One agent for everything is a vibe. Three agents with one job each is a system.”