OpenAI announced the Agents API on 10 September 2026. Available to all developers in public beta, it exposes the managed harness behind Codex through an API and supports long-running sessions, tools, MCP servers, hosted or self-hosted environments and optional subagents. Model, tool and hosted-container usage remain separately billed.
What happened: durable agent infrastructure became an API
The Agents API manages sessions, orchestration, context compaction and recovery while the developer chooses the model, tools and execution environment. An agent can work with files, run code, connect to MCP servers and produce artifacts inside an optional sandbox.
Developers can use an OpenAI-hosted Linux workspace, their own infrastructure or an integrated sandbox provider. Hosted environments can preload files, packages, skills and plugins, and their outbound network policy can be enabled, disabled or restricted to an allowlist.
The public beta also includes optional multi-agent execution. A primary agent can delegate independent work to concurrent subagents with separate contexts, while OpenAI maintains the underlying harness. The launch is versioned and based on the open-source Codex harness, but the managed service will continue changing during beta.
Why it matters: the platform boundary moved, not the accountability
Teams no longer need to build every loop for context management, tool coordination and long-session recovery. That can shorten delivery time, especially for work that spans files, browsers, code and multiple specialist tasks. It also concentrates important behaviour in a managed control plane whose beta interfaces and defaults may evolve.
A durable session can outlive the request that started it, and a subagent can multiply tool calls, data access and spend. The relevant risk is therefore not only whether a model gives a poor answer, but whether a partially completed run changes systems, leaves artifacts, repeats an action after recovery or continues consuming resources without a useful outcome.
Choosing a hosted sandbox does not define the application's security policy. The application still supplies tools and data, decides network access and handles results. Self-hosting provides more infrastructure control, but transfers isolation, patching, capacity and recovery responsibilities back to the operator.
The practical implication: write the operating contract before the prompt
Define each agent as a production identity with an approved purpose, bounded tools, explicit data sources, maximum runtime and spend, network allowlist and named owner. Separate read, draft and commit capabilities. Require fresh approval before financial, destructive, access-changing or externally visible actions, even when a session has already been approved to investigate.
Make actions idempotent and attach a stable operation identifier so retries cannot silently duplicate work. Record session, model and harness version, tool calls, approvals, subagent lineage, artifacts, cost and final disposition. Test timeouts, lost callbacks, tool failure, context compaction, sandbox termination and recovery from a partially completed workflow.
Pilot with outcome-based measures: successful completions, human corrections, unsafe or blocked actions, recovery success, elapsed time and total model, tool and container cost per accepted result. Keep a self-hosted or alternative workflow design documented for workloads that cannot tolerate public-beta change, hosted-network constraints or dependence on one managed harness.
Key takeaways
Treat the Agents API as a production control plane, with bounded identities, tools, data, runtime, networks and spend.
Design every action for safe retry and capture enough evidence to reconstruct approvals, subagent work, costs and partial failures.
Pilot against accepted business outcomes and maintain an exit path for workloads that cannot absorb public-beta or provider dependency.
Sources
This article is original analysis written by IT Web Builder. The reporting and product details below were used as references and are credited to their publishers.
OpenAI — Introducing the Agents APIPublished 10 September 2026. Primary launch announcement covering public-beta availability, managed sessions, environments, compaction, tool use, subagents and pricing structure.↗OpenAI API Documentation — Agents API overviewOfficial reference for the managed harness boundary, durable sessions, execution environments, tools, artifacts and usage charges.↗OpenAI API Documentation — OpenAI-hosted sandboxesOfficial configuration reference for files, packages, environment templates and enabled, disabled or allowlisted network access. Operational recommendations are IT Web Builder Editorial analysis.↗IT Web Builder publishes original summaries and practical commentary. We do not reproduce source articles. Product claims and announcements are linked to their original publishers so readers can verify the details.