Multi-agent workflow
Two or more AI agents interact to perform work
Core idea
A multi-agent workflow has two or more AI agent instances with distinct roles or state interacting through messages or a shared environment. This series focuses on cooperative workflows that divide a larger job and integrate the results. Separate contexts and a central orchestrator are common designs, not requirements; asking one agent to perform several roles in sequence is different.
In plain language
It resembles teachers sharing one learning goal while one develops the activity and another develops the assessment criteria. Separate sessions can use the same model, but they may repeat correlated errors and blind spots, so a second agent is not independent quality assurance by itself.
When division helps
Multi-agent workflows can use parallel assignments, planner-executor roles, sequential delegation, or review-revision loops. Parallel division is especially useful when:
- assignments own different files or deliverables;
- each assignment's required materials are clear;
- independent tasks can proceed without each other's intermediate results; and
- the integration format and final owner are defined.
Use sequential delegation or a review-revision loop when one task needs another's result. Two agents editing the same file concurrently are likely to conflict, so ownership boundaries matter. Use one agent when duplicated effort, conflicts, integration errors, and extra token usage outweigh the value of decomposition or parallel work.
Context and cost
Each agent needs the shared goal plus its task-specific material. Copying the full conversation to every agent defeats much of the focus benefit. Removing the interface between connected tasks makes their outputs impossible to combine.
Multi-agent work can reduce context for one assignment while increasing total usage through repeated instructions, extra calls, and integration. It does not automatically reduce the overall bill.
Professional terminology
Related terms include multi-agent system, task decomposition, delegation, orchestration, handoff, and integration contract. Software work also needs file ownership, interfaces, and dependency order.
Role in this blog
This series does not divide every task among agents. It separates roles when independent work or an independent review provides clear value, after defining a shared plan and handoff format.