OPENAI · DEVELOPER PLATFORM

OpenAI Agents API enters public beta with the Codex harness

What the managed agent harness covers—long sessions, context, tools, subagents, and sandboxes—and what an enterprise still needs to control.

Source update: no separate date foundUpdated 14 September 2026Approx. 6 min read
Official OpenAI artwork for the Agents API announcement
Official OpenAI artwork for the Agents API announcementOriginal image: OpenAI
News summary

Agents API lets developers create cloud agents using the harness behind Codex by specifying a model, tools, and an environment. It supports long-running work, context compaction, tool search, and subagents. Public beta status still calls for stability and governance testing before critical production use.

Best for: Engineering teams building multistep agents for repositories, documents, or back-office systems.

Points for a team pilot

ComponentManaged capabilityEnterprise responsibility
HarnessSessions, context compaction, tool use, and subagentsTask design, success criteria, and recovery
EnvironmentOpenAI-hosted, own infrastructure, or a partnerSecrets, network egress, packages, and retention
ToolsLarge tool sets with on-demand discoveryAllowlists, least privilege, and audit

What changed

OpenAI exposed an agent harness that maintains context and long-running work through an API, reducing the amount of orchestration teams need to build themselves and enabling subagents within a session.

Availability and limits

The announcement describes a public beta for all developers. There is no separate Agents API fee, but selected models and tools are metered. Teams should confirm quotas, data residency, logging, sandbox behavior, and beta terms in their own account before production.

Implications for Thai organizations

It can speed up back-office prototypes such as document checks or code changes. Do not let an agent write to production or send external data without approval. Tool permissions, secrets, and repeated-run cost are the main operational controls.

Dates and evidence

OpenAI announced and opened the public beta on September 10, 2026. No separate source-update date was found.

Before getting started

  • Choose a sandbox and network policy
  • Keep secrets out of prompts and scope tools
  • Set time, budget, and subagent limits
  • Record traces and approvals
  • Test rollback and stalled-agent recovery

Frequently asked questions

Does it replace a workflow engine?

Not entirely. It provides the harness, while the application still needs state, approvals, retries, and verifiable outcomes.

Is it production-ready?

It is a public beta. Start with bounded work and verify current terms before mission-critical use.

Continue with practical guidance for evaluation and procurement.

GPT-6 AstraAI procurement checklist
OFFICIAL & REFERENCE SOURCES

Sources used for verification

Vendor facts are separated from TechTouch guidance. Features, pricing, and commercial terms may change, so confirm the latest information at the time of purchase.

01OpenAI — Introducing the Agents API