AILet’s talk
Back to AI updates
OpenAINew features

Introducing the Agents API

Mainly relevant if your organisation builds its own AI tools or integrations rather than just using ChatGPT day-to-day — it gives developers a ready-made, managed foundation for building reliable agents that run tasks over time, instead of having to build that orchestration and sandboxing infrastructure themselves.

Read the original source 2 min read
Diagram showing an application sending tasks to the Agents API, a managed Codex harness, which sends tool calls to a sandbox and returns tool results and output back to the application.
OpenAI — Agents API architecture ↗

What changed

Source summary

OpenAI released the Agents API in public beta, giving developers a single API to build production-ready agents. It brings the same infrastructure that powers Codex — session management, tool orchestration, hosted or self-managed sandboxes, and context recovery — to external developers.

Why it matters to your work

Practical interpretation

Mainly relevant if your organisation builds its own AI tools or integrations rather than just using ChatGPT day-to-day — it gives developers a ready-made, managed foundation for building reliable agents that run tasks over time, instead of having to build that orchestration and sandboxing infrastructure themselves.

When this could help

Building agents that run long tasks, need to split work across parallel subagents, or need sandboxed code execution without building that infrastructure in-house

Brings the same managed infrastructure that powers Codex — session management, tool orchestration, hosted or self-managed sandboxes, and automatic context compaction — to any developer through a single API, with versioned harness updates on each model launch

How to get started

  1. Start with the quickstart guide to create a first agent

    Provider guidance ↗

  2. Choose an environment: the OpenAI-hosted sandbox for fast setup, or a self-hosted/partner sandbox (Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, Vercel) for more control

    Provider guidance ↗

What to check

  • Public beta — features and behaviour may change before general availability Source ↗
  • No specific session-duration or concurrency limits are documented in the announcement Source ↗
  • Sandbox choice affects cold-start time, performance, and cost, so the environment needs to be matched to the workload Source ↗

Sources & provider guides

Official documentation checked 12 Oct 2026. SCALA prompts and workflows are suggestions, not claims of hands-on testing.

Source & availability

This is a summary of a published provider announcement, with practical context from SCALA’s feed.

Plan, region, workspace and rollout conditions may affect access in Australia. Check the original source for current availability.

openai.com

This update is primarily for technical teams.