Agent orchestration

The OpenAI Agents SDK describes two patterns, and says they can be mixed. In the first, an agent is given an open-ended task and the model plans how to tackle it, using tools to take actions and acquire data, and using handoffs to delegate tasks to sub-agents¹. One form is a manager that keeps control of the conversation and calls specialist agents as tools. Another is a triage agent that routes the turn to a specialist, who then remains the active agent for the rest of the turn¹.

In the second pattern, code chooses the route. The SDK says this is more deterministic and more predictable for speed, cost and performance¹. Code can pass one agent’s output in as the next agent’s input, run an evaluator in a loop until the evaluator says the output passes, or run agents that do not depend on each other at the same time¹. One chaining example decomposes a blog post into research, an outline, the post, a critique and an improvement¹.

Anthropic draws a related line. A workflow orchestrates models and tools through predefined code paths. An agent is a system in which the model directs its own process and its own tool use². Agentic systems often trade latency and cost for better task performance, and Anthropic says to add that complexity only when it demonstrably improves the outcome². In its orchestrator-workers pattern, a central model splits a task, delegates the parts and combines the results. The examples it gives are coding work that changes several files, and search that draws on several sources².