01
Overview
OpenAI's Agents API includes native multi-agent support, allowing a main agent to break a complex task into independent pieces and delegate them to subagents that can work in parallel.
02
How subagents work
OpenAI says each subagent maintains its own context while the main agent coordinates assignments and combines results. This is useful because one branch's working notes do not have to crowd every other branch's context.
The launch example shows multi-agent mode enabled with a maximum number of concurrent subagents. Exact defaults and limits are beta details, so verify the current docs before shipping.
03
When parallel agents help
Parallel agents are strongest when subtasks are independent and can be evaluated separately. Examples include investigating multiple services during an incident, researching several markets, reviewing separate modules in a repository, or gathering evidence from different data sources.
If tasks are tightly sequential, parallelism may create more coordination work than value.
04
Design the parent agent as a coordinator
The parent agent should define the objective, split work into bounded assignments, specify evidence requirements, enforce permissions and synthesize results.
A good delegation gives each subagent a narrow job and a concrete output. 'Check deployment changes in the last hour and return evidence' is better than 'investigate everything.'
05
Keep contexts intentionally separate
Separate contexts help focus, but important shared constraints must be explicitly included in each assignment. Do not assume every subagent automatically knows every decision made elsewhere.
Use durable files or application state when multiple branches need shared facts that must remain consistent.
06
Tool permissions
Not every subagent needs every tool. A research subagent may only require web/search access, while a deployment subagent might need infrastructure APIs. Narrow tool access reduces both risk and context complexity.
07
Cost and latency
Parallel work can reduce elapsed time while increasing simultaneous model and tool usage. Measure cost per completed task rather than assuming parallel means cheaper.
Subagents may also duplicate work if decomposition is poor. Add clear scopes and stop conditions.
08
Failure handling
Plan for partial success. One subagent can fail while others complete. The coordinator should know whether to retry, continue with missing evidence, request human review or stop the workflow.
09
Good multi-agent patterns
Research fan-out: several agents gather evidence, one synthesizes. Code review split: agents inspect different modules, one consolidates findings. Incident investigation: separate agents inspect deployments, errors and dependencies. Data workflow: agents process independent partitions, one validates aggregate results.
10
Bad multi-agent patterns
Sequential tasks disguised as parallel work. Multiple agents mutating the same resource without coordination. Broad duplicated assignments with no ownership boundaries. High-risk actions delegated without approval or rollback rules.
11
Production checklist
- Verify current concurrency limits.
- Set narrow subagent scopes.
- Give only required tools.
- Define shared facts explicitly.
- Persist critical state outside model context.
- Track cost per branch and total job.
- Handle partial failures and retries.
- Require approval for high-impact mutations.
12
Bottom line
Native subagents make the Agents API more than a single long-running agent loop. The feature can shorten workflows that naturally decompose, but the gain comes from good task architecture, not from multiplying agents by default.
Sources
Primary and supporting sources
Facts were rechecked against the linked sources immediately before publication. Pricing, product availability and rollout status can change.