ARTICLE / SEPTEMBER 12, 2026
The OpenAI Agents API for Teams: What Changed on 2026-09-10 and What Did Not
OpenAI's Agents API entered public beta on 2026-09-10; here is what changes for teams, what stays unchanged, and when to adopt it.
By Kortix Blog · Published: September 12, 2026 · Last updated: September 12, 2026
What does the OpenAI Agents API change for teams?
The OpenAI Agents API is a managed, single-call layer that runs OpenAI's open-source Codex harness for you, and it entered public beta on 2026-09-10. For a team, the practical change is that you no longer have to build and maintain the harness that coordinates model calls, tools, and context: you describe the task, model, tools, and environment in one call, and OpenAI hosts and maintains the harness. (openai.com, 2026-09-10)
What stays unchanged is just as important for planning. The Responses API and the Agents SDK remain OpenAI's lower-level, code-first primitives, and OpenAI did not announce a separate Agents API fee: you continue to pay for the tokens and tools your agents use. Teams already running the Responses API or the Agents SDK are not being migrated or deprecated off them by this announcement. (openai.com, 2025-03-11, openai.com, 2026-09-10)
What's new vs what stays unchanged
| Capability or concern | Responses API + Agents SDK | Agents API (public beta) | What it means for a team |
|---|---|---|---|
| Harness ownership | You build and maintain orchestration | OpenAI hosts and maintains the Codex harness | Removes harness maintenance from your backlog |
| Creation effort | Custom orchestration code | One API call specifying task, model, tools, environment | Faster path to a production-ready agent |
| Compute environment | Your runtime and sandbox provider | OpenAI-managed sandbox, your infrastructure, or a partner | Choose isolation, VPC, and cost profile per workload |
| Multi-agent work | Handoffs you orchestrate yourself | Built-in delegation to parallel subagents | Parallel research and coding without custom orchestration |
| Long-running sessions | You implement compaction | Automatic context compaction near the limit | Workflows can span multiple context windows |
| Tool handling | You wire each tool | Tool search and programmatic tool calling | Lower token usage and a preserved model cache |
| Pricing | Tokens and tools | No additional Agents API fee; tokens and tools | No new line item to budget |
| Primitive status | The lower-level primitives | A managed layer above them | Both remain available; choose per task |
What OpenAI announced on 2026-09-10
On 2026-09-10, OpenAI introduced the Agents API in public beta, describing it as "that same harness and infrastructure that powers Codex" made available "through a simple, flexible API." The announcement states that a developer can "create a production-ready agent in a single API call by specifying the task, model, tools, and environment," with OpenAI hosting and maintaining the harness. (openai.com, 2026-09-10)
Three specifics matter for planning. First, compute is a choice, not a requirement: an agent can run "in an OpenAI-managed sandbox, on your own infrastructure, or with one of our sandbox partners," and OpenAI named Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel as first-class integration partners. Second, the Agents API is "powered by the open-source Codex harness," so the core logic that coordinates model calls, tools, and context is publicly inspectable even though OpenAI operates it. Third, the API is "available in public beta today to all developers," with "no additional fees for using the Agents API." (openai.com, 2026-09-10)
What changes for teams
A hosted harness instead of self-managed orchestration
The largest change for a team is ownership of the orchestration layer. The Codex harness is the code that coordinates model calls, tools, and context, and with the Agents API OpenAI runs and maintains it while developers can inspect the public codebase. (openai.com, 2026-09-10) The trade-off is explicit: you give up direct control of the loop in exchange for not maintaining it. OpenAI documents that the Agents SDK route keeps deployment, storage, approvals, and runtime integration inside your own application. (OpenAI docs, accessed 2026-09-12)
Compute and sandbox choice: OpenAI-managed, your infrastructure, or a partner
The Agents API does not force a single hosting model. OpenAI offers a hosted sandbox that "leverages the same sandboxing infrastructure that powers Codex and ChatGPT," and it also supports a self-hosted environment or a partner sandbox. (openai.com, 2026-09-10) The enterprise channel already exists: OpenAI states that enterprises "can now deploy agents built on Codex harness to Cloudflare," where "the Codex harness is now generally available in Cloudflare Sandboxes." (openai.com, 2026-04-13)
Multi-agent delegation and parallel subagents
With multi-agent support, the Agents API "can break complex tasks into independent pieces and delegate them to subagents that work in parallel." Each subagent keeps its own context while the main agent coordinates the work and combines the results. For a team, that is orchestration logic you would otherwise write, test, and maintain yourself. (openai.com, 2026-09-10)
Long-running sessions and automatic context compaction
The Agents API "automatically compacts earlier context as a session approaches its context limit," which lets developers "build workflows that span multiple context windows without implementing their own compaction logic." If your team has written or bought context-summarization code, this moves that concern into the managed harness. (openai.com, 2026-09-10)
Tool search and programmatic tool calling: token and cache effects
Two features target cost and latency. Tool search "loads relevant tool definitions as needed, helping reduce token usage and cost while preserving the model's cache," and programmatic tool calling lets agents "run calls in parallel, chain related operations, and filter or combine results in code" so only relevant results return to context. Both change the arithmetic of running a large tool catalog. (openai.com, 2026-09-10)
What stays unchanged
The Responses API and Agents SDK remain the lower-level primitives
OpenAI introduced the Responses API and the Agents SDK on 2025-03-11 as "building blocks" for agents: the Responses API combined "the simplicity of the Chat Completions API with the tool use capabilities of the Assistants API," and the Agents SDK was released "to orchestrate single-agent and multi-agent workflows." Those primitives remain available and documented. (openai.com, 2025-03-11) OpenAI's current docs still list all three runtimes side by side and describe the Agents API as the managed option that "runs the Codex harness and manages the underlying agent infrastructure." (OpenAI docs, accessed 2026-09-12)
Code-first agent building
The Agents API is itself code-first: you create sessions and configure agents, tools, and environments through an API, not a visual canvas. (openai.com, 2025-10-06)
Pricing: tokens and tools, no separate Agents API fee
There is no new subscription or platform fee to evaluate. The announcement says there are "no additional fees for using the Agents API – you simply pay for the tokens and tools your agents use," and the docs add that OpenAI-hosted sandboxes use standard container rates. Your existing token and tool budget model still applies. (openai.com, 2026-09-10, OpenAI docs, accessed 2026-09-12)
Agents API vs Responses API vs Agents SDK: choose the right layer
The three runtimes answer different questions, and OpenAI's own comparison is the clearest way to choose. The Agents API is for "long-running tasks where OpenAI manages the agent and saves its progress"; the Agents SDK is for "building agents with custom tools and workflows in your application"; the Responses API is for "calling models directly or building an agent from scratch." (OpenAI docs, accessed 2026-09-12)
| Dimension | Agents API | Agents SDK | Responses API |
|---|---|---|---|
| Use for | Long-running tasks where OpenAI manages the agent and saves progress | Custom tools and workflows inside your application | Calling models directly or building an agent from scratch |
| Where the agent runs | OpenAI-managed Codex harness | The SDK runs in your application | Your application, with optional hosted orchestration |
| Integration effort | Low | Medium | High |
| State between tasks | Saved session configuration, turns, and items | Your storage and SDK sessions, or Responses conversation state | Manual history, response chaining, or Conversations |
| Tool execution | Service-connected tools, application handlers, optional sandbox | Tools and integrations you configure | Hosted tools and tools your application runs |
| Execution environment | OpenAI hosted sandbox, self-hosted sandbox, or none | Your runtime and sandbox provider integrations | Your own execution environment |
Source: OpenAI's agent runtime comparison. (OpenAI docs, accessed 2026-09-12)
A practical rule follows from the table: choose the Agents API for a managed, long-running agent when you are comfortable with OpenAI running the loop; choose the Agents SDK when orchestration must stay in your process; stay on the Responses API for a narrower integration that does not need an agent loop. (OpenAI docs, accessed 2026-09-12)
When to adopt the Agents API now vs wait
Adopt now if the work is long-running, tool-heavy, and benefits from parallelism. Research, analysis, and coding are the cases OpenAI names for multi-agent delegation, and the API is in public beta to all developers at no additional fee. (openai.com, 2026-09-10)
Wait, or pilot behind a flag, if any of these apply. It is a public beta, and OpenAI says it will "iterate quickly based on your feedback as we work toward general availability," so interfaces can move. Data residency and retention are constraints: the docs state that the Agents API "currently supports data residency only in the United States and does not support Zero Data Retention (ZDR)," and that choosing a self-hosted sandbox does not make it ZDR-eligible. (OpenAI docs, accessed 2026-09-12) Teams with EU residency or ZDR requirements should treat those as blockers until OpenAI changes them.
A reasonable evaluation is a bounded pilot on one long-running workflow, measured against the same task on your current stack. Compare task-success rate, cost per task, and time-to-first-working-agent, and keep the Responses API or Agents SDK path available so the decision stays reversible. (OpenAI docs, accessed 2026-09-12)
How Kortix relates to this decision
Kortix is an open-source AI command center that keeps a company's agents, skills, and memory in one versioned repository the team owns, and it is self-hostable with any model and your own keys. (kortix.com) Its official repository, kortix-ai/suna, is described as "The open-source AI Management System." (GitHub)
That places Kortix at a different layer from the Agents API: the Agents API is a managed runtime where OpenAI operates the harness, while Kortix is a self-hostable platform for running and governing a team's agents across sessions. For a team weighing a hosted harness against infrastructure control, the two are complementary rather than either-or, and the engineering work of making long-running workflows reliable applies either way. (Kortix Blog)
Background: OpenAI's consolidation path
This section is background only, not part of the 2026-09-10 announcement. OpenAI's AgentKit launch on 2025-10-06 included a visual Agent Builder and Evals, but the page now carries an update: on 2026-06-03 OpenAI said it is "winding down the Agent Builder and Evals products," that from November 30, 2026 they "will no longer be available on the OpenAI platform," and that for workflows that should continue as code it recommends the Agents SDK. (openai.com, 2025-10-06) Read together with the Agents API launch, the direction is a code-first stack — the Agents SDK plus the managed Agents API — rather than a visual builder.
Frequently asked questions
Is the OpenAI Agents API generally available?
No. It entered public beta on 2026-09-10 and is available to all developers, and OpenAI says it will iterate quickly toward general availability. That beta status is the main reason to pilot before standardizing. (openai.com, 2026-09-10)
Do we have to migrate off the Responses API or Agents SDK?
No. Both remain available lower-level primitives, and OpenAI documents them alongside the Agents API as different runtime choices rather than as a replacement path. (OpenAI docs, accessed 2026-09-12)
How much does the OpenAI Agents API cost?
There is no separate Agents API fee. OpenAI states that you "simply pay for the tokens and tools your agents use," and the docs add that OpenAI-hosted sandboxes use standard container rates. (openai.com, 2026-09-10, OpenAI docs, accessed 2026-09-12)
What is the open-source Codex harness?
The Codex harness is the core logic that coordinates model calls, tools, and context, and the Agents API is powered by its open-source version. OpenAI operates and maintains the harness while developers can inspect the public codebase. (openai.com, 2026-09-10)
Can we run the Agents API on our own infrastructure?
Yes. OpenAI offers an OpenAI-managed sandbox, a self-hosted environment on your infrastructure, and partner sandboxes including Blaxel, Cloudflare, Daytona, DigitalOcean, E2B, Modal, Oracle, Runloop, and Vercel. (openai.com, 2026-09-10)
What is automatic context compaction?
Automatic context compaction is the managed harness summarizing earlier context as a session approaches its context limit, so a workflow can continue across multiple context windows without your team implementing its own compaction logic. It is one of the maintenance tasks the Agents API removes from your codebase. (openai.com, 2026-09-10)
Does the Agents API support multi-agent workflows?
Yes. Multi-agent support lets the API break complex tasks into independent pieces and delegate them to subagents that work in parallel, each with its own context while the main agent coordinates the results. (openai.com, 2026-09-10)
Should a team adopt the Agents API during public beta?
Adopt or pilot it if your workload is long-running and tool-heavy, you accept that beta interfaces can move, and United States data residency is acceptable. Wait if you require EU residency or Zero Data Retention, because the docs state the Agents API supports US residency only and is not ZDR-eligible, and a self-hosted sandbox does not change that. (OpenAI docs, accessed 2026-09-12)
Sources and last verified date
All primary claims in this article were verified against the sources below on 2026-09-12.
- OpenAI, "Introducing the Agents API," 2026-09-10 — https://openai.com/index/introducing-the-agents-api
- OpenAI, "New tools for building agents," 2025-03-11 — https://openai.com/index/new-tools-for-building-agents
- OpenAI, "Enterprises power agentic workflows in Cloudflare Agent Cloud with OpenAI," 2026-04-13 — https://openai.com/index/cloudflare-openai-agent-cloud
- OpenAI, "Introducing AgentKit," 2025-10-06, with the 2026-06-03 wind-down update — https://openai.com/index/introducing-agentkit/
- OpenAI developer docs, "Agents API overview" and the agent runtime comparison, accessed 2026-09-12 — https://developers.openai.com/api/docs/guides/agents-api/overview and https://developers.openai.com/api/docs/guides/agents
- Kortix, official site — https://kortix.com
- Kortix, official GitHub repository — https://github.com/kortix-ai/suna
Related reading on this blog: How to build reliable AI agent workflows maps each production failure mode to a concrete control and the signal that proves it works, and Welcome to the Kortix Blog explains what this blog covers.