iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
Use subagents when a larger task can be split into bounded, independent pieces whose results a coordinator can combine. Keep short tasks and steps that depend closely on one another with a single agent. “Agent teams” is a useful informal label for coordinated agents, not one standardized runtime: it may describe API delegation or a person supervising multiple coding agents in the Codex app, and those are different implementations.
What is the difference between subagents and agent teams?
In OpenAI’s API guidance, a root or coordinating agent delegates work to subagents. Each subagent has its own context, and independent tasks can run in parallel; the coordinator then synthesizes their results. OpenAI recommends keeping short tasks and dependent steps in the main agent (OpenAI Agents API multi-agent guide).
“Agent team” is not a single, standardized runtime name across the documentation discussed here. In the Codex app, it can mean a person managing multiple agent threads and reviewing their work. In API workflows, it can mean a coordinator delegating to subagents. The concepts overlap, but they do not imply the same orchestration, concurrency, isolation, or billing model. The Codex app announcement describes its app workflow; API delegation is covered separately in the Agents API guide and Responses multi-agent guide.
When should I use subagents?
Delegate when the work divides into meaningful, independent assignments, each can be described with a bounded deliverable, and a coordinator can reconcile the results. Parallel execution or focused context should be worth the added orchestration and review. OpenAI’s examples include independent document review, comparing release notes, and investigating separate possible causes (Agents API guide; Responses guide).
- Good candidate: Ask separate agents to review different documents for a defined set of risks, then have the coordinator compare findings.
- Good candidate: Investigate distinct hypotheses about a bug in parallel when each investigation can return evidence and a conclusion.
- Poor candidate: Split a brief, sequential task into agents when each step needs the previous step’s result.
Write bounded delegation prompts
For each subagent, specify the task, its boundaries, expected output, and the files or sources it may use. Have the coordinator check the outputs, resolve disagreements, and combine them into one answer. Delegation does not itself guarantee correctness; synthesis and review remain part of the workflow.
When should I keep the work in one agent?
Keep the task with the main agent when it is short, when steps depend on prior results, or when handing work off and reconciling it would cost more than any parallel progress saves. In a tightly coupled coding change, for example, a sequence of edits and checks may be easier to manage as one continuous task than as independently assigned pieces. This follows OpenAI’s guidance to keep short tasks and dependent steps in the main agent (Agents API guide).
How do I choose an implementation?
Once delegation makes sense, choose a surface based on who owns orchestration, how state is handled, how much integration work your application can take on, and where tools and code run. OpenAI’s agent runtime guide compares agent use, execution location, integration effort, state, tools, and environment.
Recommended Free Tools
| Option | Orchestration and state | Good fit | Trade-off |
|---|---|---|---|
| Agents API | OpenAI manages the Codex harness, session, orchestration, context compaction, and recovery; the application supplies the task, tools, and configuration. | Long-running managed workflows where reducing integration effort and using managed session state matter. | Less runtime ownership than an SDK you operate. Check current beta status and usage costs (Agents API overview; multi-agent guide). |
| Agents SDK | The application uses the SDK runner and controls deployment, storage, approvals, and runtime integration. | Reusable custom workflows built around your own tools and application logic. | More integration work and responsibility for state and runtime choices (agent runtime guide). |
| Responses API | The application works more directly with model responses and can build orchestration itself or use available hosted orchestration features. | Direct model access or custom integration where the developer wants control over the agent loop. | More application responsibility; the reviewed multi-agent feature is described as beta (Responses multi-agent guide; agent runtime guide). |
| Codex app | A person manages agent threads and reviews changes. Built-in worktrees provide isolated copies of a repository for agent work. | Parallel coding tasks where human review stays in the workflow. | An app workflow is not the same as API subagent orchestration; confirm current availability and plan limits (Codex app announcement; Codex plan support page). |
What concurrency settings and availability limits apply?
Concurrency values are specific to the API surface, not universal “agent team” limits. The Agents API multi-agent guide says to enable multi-agent orchestration when creating a session and configure max_concurrent_subagents; it describes a configurable default of six when enabled. The Responses multi-agent guide instead describes max_concurrent_subagent_turns, which limits active subagent turns across the tree, with a configurable default of three. These settings are not interchangeable (Agents API guide; Responses guide).
The reviewed documentation labels the Responses multi-agent feature and Agents API beta. Model eligibility, defaults, beta status, and plan access can change, so check the current documentation and your account before committing a production design (Responses guide; Agents API overview).
Are agent teams faster or more accurate than one agent?
There is no general controlled comparison in the cited OpenAI documentation establishing that multiple agents are always faster, cheaper, or more accurate than one. A parallel workflow may save time when assignments are genuinely independent, but orchestration, handoffs, and review add work. Whether the trade-off pays off depends on the task and the implementation (Agents API guide; Responses guide; agent runtime guide).
If performance matters, compare the same representative workload using one agent and your proposed delegated workflow. Track end-to-end completion time, review effort, and whether the final result meets your quality criteria; do not infer a general speed or quality gain from concurrency alone.
How should I estimate the cost?
There is no universal price for an agent team. OpenAI’s Agents API overview says model usage is billed at the selected model’s API rates, OpenAI tools use their standard rates, and OpenAI-hosted sandboxes use standard container rates. Estimate the model, tool, and sandbox usage your workflow is likely to incur, using the applicable rates and configuration (Agents API overview).
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
