Orchestration becomes infrastructure for long-running parallel agents
Source → Observed → Interpretation → Model implication
Introducing the Agents API
View source →OpenAI says useful long-running agents need a harness that manages context, uses tools efficiently, coordinates subagents, preserves intermediate work, and keeps execution reliable for days.
The Agents API separates the managed Codex harness from the agent's compute environment, which may be an OpenAI-hosted sandbox, customer infrastructure, or an integrated sandbox provider.
OpenAI reports that the harness compacts earlier context for sessions spanning multiple context windows, loads relevant tool definitions on demand, and lets agents run tool calls in parallel or combine results before returning relevant information to model context.
OpenAI reports that multi-agent support delegates independent work to parallel subagents with separate contexts while a main agent coordinates and combines their results.
Ciridae reports that after adopting the Agents API its evaluation score rose from 0.71 to 0.85 and its previous difficulty observing and orchestrating subagents was reduced, with a 4x latency improvement.
Customer reports published by OpenAI include asynchronous fan-out across hundreds of agents and thousands of long-running agents coordinating work across production logistics workflows lasting hours or days.
The source shows context management, durable execution, tool mediation, and subagent coordination moving from application-specific implementation into a reusable platform layer. This refines Selection by identifying environmental capabilities that support longer-running and more parallel configurations. The Ciridae comparison also supplies direct, though first-party and bundled, evidence that a cooperating subagent configuration with dedicated orchestration can improve latency and evaluated quality over its prior setup. The separation of harness, sandbox, tools, and coordinated subagents resembles differentiated capabilities, but the source does not establish the persistent integrated higher-order engineering object required by Specialization.
SUPPORTS. Supports Cooperation with a reported before-and-after outcome: dedicated subagent orchestration improved both latency and an evaluation score compared with a customer's previous setup. More broadly, the API suggests that as agents become long-running and parallel, context management and coordination become reusable infrastructure rather than remaining embedded in each agent application. The evidence is not sufficient to establish persistent Specialization.
What this does not establish
- The Ciridae result does not isolate subagent cooperation from the other harness and API changes introduced during migration.
- The customer outcomes are first-party testimonials published by OpenAI and do not provide task definitions, evaluation methodology, repeated-trial results, or independent replication.
- Reports of hundreds or thousands of agents come from different customer workloads and do not establish equivalent scale for coding-agent work.
- The source does not compare persistent cooperating and isolated configurations across repeated software-engineering tasks.
- Distinct harness, sandbox, tool, and subagent roles do not by themselves establish stable Specialization or a meaningful higher-order engineering object.
Across repeated software-engineering tasks, which dedicated orchestration capabilities preserve the reported quality and latency advantage as subagent concurrency and task duration increase, and at what coordination cost?