Workflow, Agent, or Multi-Agent

A single prompt with tools is already a system. Beyond that it is a question of architecture, and the word "agent" mustn't stand in for the decision. A workflow follows a path set by code. An agent picks the next steps itself. Multi-agent adds several independent contexts and, with them, the problem of coordination.

The choice comes down conveniently to four options, each with its own price.

• Single call - minimal latency and state, but limited complexity. For extraction, classification, a grounded answer.
• Deterministic workflow - predictable order at the price of describing the branches in code. For a support flow, a pipeline, approvals.
• Single agent - flexibility on intermediate results at the price of extra tokens, traces, and failure modes. For diagnosis, research, open-ended tools.
• Multi-agent - isolated contexts, specialization, and parallelism at the price of coordination, duplication, and high cost. For independent expensive workstreams or sequential handoffs with different contracts.

Before taking multi-agent, answer four questions about the task honestly - the easiest way is to run them through the architecture choice.

And two sober reminders.

Multi-agent isn't free speedup. For its own research system Anthropic reported substantially higher token usage and a poor fit for tasks with tight dependencies. This is an observation about a specific system, not a universal coefficient.

Check the feature's status. As of the audit date the built-in Multi-agent in OpenAI Responses is marked beta and aimed primarily at purely parallel, independent workstreams. A handoff or agent-as-tool architecture can be assembled differently and not tied to this beta.