
OpenAI’s Agents API public beta gives developers a way to build cloud-based agents that run for hours, execute code, process files and use tools without operating the underlying systems themselves. The API runs on the same infrastructure that powers Codex and ChatGPT, according to The Decoder.
The public beta launched on September 10, 2026, according to the OpenAI Agents API integration status reviewed by EvoLink on October 2, 2026. The September 11 report from The Decoder described billing as based solely on token usage.
What the OpenAI Agents API Public Beta Includes
The API provides a set of building blocks for agents that complete connected tasks. Supported capabilities include automatic context management, parallel tool use, delegation to sub-agents, code execution and file processing. The API supports MCP, custom functions and built-in tools such as web search.
Developers can select an execution environment for the work. The available options include OpenAI-hosted sandboxes and execution environments supplied by Cloudflare, Vercel and Oracle. The API builds on the open-source Codex harness.
The OpenAI Agents SDK documentation describes agents as language models equipped with instructions and tools. It also includes handoffs for delegating work to other agents, guardrails for validating inputs and outputs, tracing for debugging workflows, persistent sessions and mechanisms for involving a person during an agent run. OpenAI’s Agents SDK documentation says the SDK uses the Responses API by default when the managed runtime needs to handle turns, tools, guardrails, handoffs or sessions.
Which Capability Fits a Business Workflow?
| Agents API capability | Concrete business workflow | Visible output | Review point |
|---|---|---|---|
| Web search and parallel tool use | Gather market or supplier information from several connected sources | A sourced research brief or source list | Verify source claims, dates and citations before publication or use |
| Code execution and file processing | Inspect repository files, apply changes and run targeted tests | A patch and test results | Review the diff, test outcome and authorization before code is merged |
| Sub-agents and context management | Split independent review questions across specialist agents and combine their findings | A consolidated report with supporting material | Check that the combined report preserves source context and unresolved issues |
| Artifact SDK and document skills | Process source files and produce artifacts your application can retrieve | Files that become deliverables | Require approval of edited documents and check calculations, formatting and final content |
| Persistent sessions and scheduled automations | Run a follow-up task after an initial request or repeat a defined task on a schedule | A completed follow-up result or recurring task status | Define which follow-ups may run automatically and who receives the result |
| Tracing and execution evidence | Review a multi-step run after a stalled task or failed tool call | A trace that can be linked to the business job | Reconcile tool calls and session activity with the final deliverable |
EvoLink’s integration review, published October 2, 2026, describes a file-reconciliation use case in which an agent processes CSV exports and produces a workbook containing totals, exceptions and records for review. Its integration was still in progress and was not open for calls at the time of that review. EvoLink’s integration status also notes that the application must authorize business tools and check the final result.
What Changed in the Hosted Sandbox
OpenAI’s changelog records several updates to the public beta.
OpenAI also added scheduled automations to the Agents API public beta. Sessions can schedule one-time or recurring tasks, including follow-ups in the current session. OpenAI’s developer changelog describes this as an automations feature for new or existing sessions.
Beta Changes to Track
- October 23, 2026: OpenAI added scheduled automations to the Agents API public beta. Sessions can schedule one-time or recurring tasks, including follow-ups in the current session. OpenAI’s changelog describes this as an automations feature for new or existing sessions.
- October 30, 2026: OpenAI enabled updates to the tool list of an existing session. Updates replace the full tool list, with function tools and MCP servers changing on the next model call and built-in tools changing by the next turn. OpenAI’s changelog also records background context compaction by default and notes that it may increase token usage.
- October 30, 2026: OpenAI added automatic review for eligible projects. The documented option asks a model to approve or deny supported tool and permission requests, with additional token billing. OpenAI’s changelog lists limitations for supported requests.
How to Review Billing for Hosted Sandboxes
The Decoder’s September 11, 2026 report said there were no extra fees and that billing was based solely on token usage. OpenAI’s October 30 update warns that background context compaction may increase token usage. Review the relevant billing documentation before using the service.
Token usage remains relevant to the overall cost. The changelog provides product changes, but the reviewed material does not publish a universal dollar rate for active-turn sandbox compute or state fixed rate limits for the public beta.
A Practical Review Checklist Before You Let an Agent Act
OpenAI’s Agents SDK documentation says guardrails can validate agent inputs and outputs, while its human-in-the-loop mechanisms can involve people during a run. A business owner can ask a developer to verify the following before allowing an agent to act on company information:
- Review the permissions and credentials granted for each business system.
- Check how source claims, URLs and file changes are recorded.
- Ask how a failed multi-hour run resumes or reports incomplete work.
- Require a reconciliation of tool calls, active turns and token usage with the invoice.
- Assign an owner to approve final research, document and code outputs.
The SDK documentation also identifies tracing as a way to visualize, debug and monitor agentic workflows. That record can help a business connect an agent run to a specific job and review what happened when a task stalls.
What Is Confirmed Versus Not yet Documented for the Beta?
Confirmed
- The public beta began on September 10, 2026, according to EvoLink’s October 2 review.
- The API supports long-running sessions, automatic context management, parallel tool use, sub-agents, code execution and file processing.
- Scheduled automations support one-time and recurring tasks.
- Existing sessions can receive updated tool lists.
Not Documented in the Reviewed Material
- Whether there is a fixed public-beta rate limit for each API feature.
- Whether there is a universal dollar price for hosted sandbox compute during active turns.
What to Do Now
- Choose one workflow with a defined final output.
- List the files, sources and business tools the workflow needs.
- Assign a reviewer for research claims, document changes and code changes.
- Test token usage and active-turn compute separately.
- Review the session trace and invoice after the test run.
FAQ
What Is the OpenAI Agents API?
The OpenAI Agents API is a public beta interface that lets developers build cloud-based agents capable of running for hours, executing code and processing files. It runs on the same infrastructure that powers Codex and ChatGPT.
What Features Does the Agents API Include?
The API includes automatic context management, parallel tool use and the ability to delegate tasks to sub-agents. It supports MCP, custom functions and built-in tools such as web search.
Where Can Agents Built With the API Run?
Agents can run in OpenAI-hosted sandboxes or on infrastructure from partners including Cloudflare, Vercel and Oracle. The Decoder reported on September 11, 2026, that billing was based solely on token usage.
Related Coverage
- OpenAI’s isolated AI agents built hidden CPU cache channels across sandboxes (October 26, 2026): Business owners running AI agents should consider reviewing whether sandbox boundaries also separate agents at the physical-host level.
- OpenAI releases GPT-6 Astra, its most intelligent and most aligned model (September 4, 2026): GPT-6 Astra gives developers an agent that can finish multi-step digital work faster while staying within the boundaries set for it, according to OpenAI’s published results.
- Manus Cloud Computer: Autonomous AI Agents That Work 24/7 in the Cloud – BizScoreAI
Sources
See our editorial disclaimer for how our articles are produced.
Run a free scan to see your AI Visibility Score, SEO rating, and local citation accuracy.