Harness-as-a-Service: Building with the Agents API | DevDay 2026
Use the Agents API to offload long-running agent infrastructure while you define the tools, data, environment, and human approval points.
- Choose a hosted or your own sandbox, then stream session events to show users the agent’s progress.
- Use skills, tools, and limited subagents to structure work; higher subagent concurrency may use more tokens.
- Keep consequential actions behind human approval, and request output files that make the agent’s reasoning checkable.
The flight-rebooking demo shows session setup, live activity, subagents, files, and human approval. · Best for: Developers building products with long-running or parallel agents.
OpenAI presents the Agents API as a managed service built on the Codex harness. Developers specify tools, data, instructions, and the execution environment; the service manages the harness, long-running sessions, and related infrastructure.
Why offer a managed agent harness0:21
- More capable models can handle longer, more involved tasks, which require more than a basic inference API.
- The Responses API is a foundation for agentic workflows; the Agents API manages a Codex-based harness for developers.
- Teams can choose tools, data, instructions, and environment without running the whole harness themselves.
Separate the harness from its execution environment5:58
- The cloud-hosted harness can send commands to a developer-managed or OpenAI-hosted environment and receive results.
- The API supports terminal work in sandboxes; the presenters also announce computer use in OpenAI-hosted environments.
Show progress and preserve long-running work7:29
- Users need visibility into changing agent activity, especially during long tasks.
- Session turn information helps developers present tool calls and responses without rebuilding their relationships.
- Compaction carries useful context forward; service infrastructure also tracks state and resumes sessions after failures or deployments.
- If a connection drops during an action, the service must account for what happened in the execution environment before continuing.
Flight rebooking demo setup11:42
- The demo starts separate sessions for difficult rebooking cases, including a traveler with a destination constraint and a group that must travel together.
- The JavaScript example configures an OpenAI-hosted sandbox, disables network access broadly, and streams events for a live UI.
- A local skill provides rebooking steps; capability directories can expose skills and paths to plugins, including MCP servers.
- The flight inventory tool provides advisory results and can hold seats; a human must approve to confirm a rebooking.
- Two subagents check seating and arrival constraints. The demo limits concurrency to two and notes higher limits may increase token usage.
- The agent returns proposal and comparison JSON files plus a Python script for deterministic checks.
Review proposals and human approval15:07
- For the individual booking, the agent checked availability, delegated checks, and returned a recommendation with held seats for approval.
- For the family booking, the agent ruled out options that lacked enough seats or could not keep the group together.
- The UI exposes activity, workspace checks, and output files before a human approves either rebooking.
Use cases and closing17:08
- The Agents API lets developers focus on application rules and approval flows instead of maintaining the harness infrastructure.
- The presenters cite asynchronous work in Notion, long-running tasks at Rakuten, and synthetic audiences at Bain, with human experts checking results.
- Bedrock managed agents powered by OpenAI are presented as an option for builders staying in the AWS ecosystem.
- The speakers highlight scaling parallelizable work across many agents as a possible direction.
Step by step
- Create an Agents API session and select an OpenAI-hosted or supported external sandbox.
- Configure the environment, including network access, and enable event streaming if the interface needs live updates.
- Provide instructions and load relevant skills or plugins through capability directories.
- Define tools and their limits; make clear when results are advisory and which actions require human approval.
- If useful, enable multi-agent mode and set a concurrency limit suited to the work.
- Specify structured output files and checks, then monitor activity and review results before approving consequential actions.
Tools & resources mentioned
- Agents API — managed service for running agent sessions and harness infrastructure.
- Responses API — lower-level foundation described for building agentic workflows.
- Codex — harness on which the Agents API is based.
- Agents SDK — local harness mentioned as an earlier option for developers managing their own agents.
- OpenAI-hosted sandbox — execution environment used in the demo; developers can also bring their own or use a third-party provider.
- Capability directories and skills — expose agent skills and paths to plugins, including MCP servers.
- Flight inventory tool — checks available flights and may place holds; the demo requires human approval to confirm.
- Notion — example of assigning asynchronous work to agents from team workflows.
- Amazon Bedrock managed agents — OpenAI-powered option for developers building within AWS.
Notes are AI-generated from the video's transcript and may miss nuance.