Managed Agents appear to be the headline OpenAI is building towards for DevDay 2026, according to a report published on 7 September 2026 by TestingCatalog, which found evidence in the company’s own codebase for creating agents, configuring environments, and enabling skills and plugins.
The report is explicit about what it is and is not. The agents interface is not live for anyone yet. What exists are strings and structures in shipped code, which is a reliable signal of intent and an unreliable signal of timing. TestingCatalog’s Alexey Shabanov writes that OpenAI is working on an implementation that “broadly follows what Anthropic currently offers.”
That comparison is the most useful thing in the report, because Anthropic’s version is documented, in public beta and running in production. If OpenAI is following that template, the shape of the announcement is already knowable in some detail even before Sam Altman walks on stage.
This article separates the confirmed calendar from the leaked feature set, maps the leak onto Anthropic’s published architecture, and sets out the commercial reason OpenAI would ship this now. Where a claim rests on a codebase reading rather than an announcement, it says so.
The operational questions here are the same ones facing any business deploying autonomous AI agents without a human watching every step.
Table of contents
- What the Managed Agents Leak Actually Says
- DevDay 2026: Date, Venue and What Is Already Confirmed
- How Anthropic’s Managed Agents Set the Template
- Why OpenAI’s Agent Stack Left Room for Managed Agents
- Environments: The Half of Managed Agents That Decides Adoption
- Skills, Plugins and the Managed Agents Console
- The Ads Angle: When a ChatGPT Ad Resolves to Managed Agents
- Price Is the Real Question for Managed Agents
- Managed Agents Versus Rolling Your Own Loop
- What Could Go Wrong With Managed Agents
- How to Prepare Before 29 September
- Managed Agents FAQ
- References
What the Managed Agents Leak Actually Says
The evidence is narrow, specific, and worth reading precisely rather than generously.
The four capabilities found in the code
TestingCatalog reports that the codebase suggests functionality “very similar to Anthropic’s offering, including creating different agents and environments, managing them, and enabling specific skills and plugins.” Four nouns do a lot of work in that sentence: agents, environments, skills, plugins.
What is explicitly not available
The same report states that access to the agents UI “is not available yet.” No endpoint has been documented, no beta header published, no pricing disclosed. Anyone describing Managed Agents as launched is ahead of the evidence.
Why a codebase reading is still meaningful
Shipped client code carries the feature names a team has committed to internally. It rarely carries features that were abandoned months earlier. The signal is directional, and TestingCatalog has a track record on OpenAI pre-announcements, including a correct call on Agent Builder before DevDay 2025.
What else is expected on the day
Managed Agents will not be the only item. The report notes that “model updates, security solutions, and more are also expected,” which fits the pattern of every DevDay so far.
| Claim | Evidence level | Source |
|---|---|---|
| DevDay is 29 September 2026 at Fort Mason | Confirmed | OpenAI DevDay site |
| Managed Agents with environments, skills, plugins | Codebase reading | TestingCatalog |
| Ads that resolve to an agent | Reported work in progress | TestingCatalog |
| Agent Builder retires 30 November 2026 | Confirmed | OpenAI AgentKit page |
| Pricing for Managed Agents | Unknown | No source |
DevDay 2026: Date, Venue and What Is Already Confirmed
Unlike the feature leak, the event itself is fully documented on OpenAI’s own site.
The confirmed logistics
DevDay 2026 runs on Tuesday 29 September 2026 at Fort Mason in San Francisco. The opening keynote begins at 10:00 a.m. Pacific and features Sam Altman. In-person tickets cost $650 and applications have closed; selected applicants get 14 days to register before their place goes to the waitlist.
The free route in
The keynote is livestreamed and free to all. For a product announcement aimed at developers and businesses, that livestream is where the substance lands, not the $650 room.
DevDay Exchanges
OpenAI is also running DevDay Exchanges in eight cities: Bengaluru, Tokyo, Seoul, Paris, Berlin, London, São Paulo and Mexico City. The geographic spread matters for an enterprise product, because a managed runtime raises data-residency questions in every one of those markets.
| Time (PT) | Session |
|---|---|
| 8:00 a.m. | Breakfast and check-in |
| 10:00 a.m. | Opening keynote with Sam Altman (livestreamed) |
| 11:15 a.m. – 3:30 p.m. | Breakouts, technical sessions, demos and workshops |
| 4:00 p.m. | Closing session |
| 4:45 – 7:00 p.m. | Reception |
How Anthropic's Managed Agents Set the Template
If OpenAI is following Anthropic’s design, the published documentation for Claude Managed Agents is the best available preview.
Four concepts, not one product
Anthropic’s version is built on four objects. An agent is the model, system prompt, tools, MCP servers and skills. An environment is the configuration for where sessions run. A session is a running instance doing a specific task. Events are the messages passed between the application and the agent.
The pitch is subtraction
The value proposition is what you stop building. Instead of writing your own agent loop, tool execution layer and sandbox, you get a runtime where the model can read files, run commands, browse the web and execute code. Prompt caching and context compaction are handled by the harness.
It is a beta, with real caveats
Claude Managed Agents is in beta and every endpoint requires the managed-agents-2026-04-01 beta header. Because sessions are stateful by design and store history server-side, the product is not currently eligible for Zero Data Retention or HIPAA Business Associate Agreement coverage. That is a genuine constraint for regulated buyers.
The published limits
Anthropic rate-limits Managed Agents per organisation: 300 requests per minute on create endpoints such as agents, sessions and environments, and 1,200 requests per minute on read endpoints.
Why OpenAI's Agent Stack Left Room for Managed Agents
The leak makes more sense once you look at what OpenAI has already retired.
Three attempts in three years
The sequence is short and fast. Custom GPTs arrived in November 2023. AgentKit and its visual Agent Builder followed on 6 October 2025 at DevDay. Workspace Agents launched on 22 April 2026, plugging into Slack, Google Drive, Microsoft apps, Salesforce, Notion and Atlassian Rovo.
The retirement that forces a replacement
On 3 June 2026 OpenAI announced it is winding down Agent Builder and Evals. From 30 November 2026 they will no longer be available on the platform. OpenAI’s recommended successors are the Agents SDK for workflows that should stay as code, and Workspace Agents for natural-language use cases.
Why the visual builder aged so fast
TestingCatalog’s explanation is that agent capabilities “evolved enough to cover pipelines on their own, without users needing to build connections manually.” A visual canvas is a workflow automation tool, and a sufficiently capable model no longer needs one. Ramp reported getting an agent live in two sprints rather than two quarters using Agent Builder; a year later, the canvas itself is the overhead.
The gap in the middle
That leaves a hole. The Agents SDK gives you code and no runtime. Workspace Agents gives you a runtime and no code. Managed Agents would be the missing middle: your code, OpenAI’s runtime.
Environments: The Half of Managed Agents That Decides Adoption
The word “environments” in the leak is doing more work than “agents”, and it is the part enterprise buyers should read first.
What an environment actually is
In Anthropic’s model, an environment is the configuration for where a session runs: either a vendor-managed cloud sandbox, or a self-hosted sandbox on infrastructure you control. It is the boundary between “we run your code” and “you run your code, we drive it.”
Why self-hosting is the enterprise unlock
Anthropic documents self-hosted sandboxes explicitly for compliance and data-residency requirements. For a bank in Frankfurt or a hospital group in London, that option is not a nice-to-have; it is the difference between a procurement approval and a rejection. If OpenAI ships environments without a self-hosted path, the addressable market shrinks sharply.
The operational shape of a sandbox
Sandboxed execution means the model gets a real shell. Anthropic’s built-in toolset covers bash, file operations including read, write, edit, glob and grep, plus web search and fetch that can be restricted to an allowlist or blocklist of domains. Every one of those is a permission decision someone has to own.
Long-running work changes the failure modes
These runtimes are built for tasks that run for minutes or hours across many tool calls, with persistent filesystems and conversation history between interactions. A ten-second API call that fails is an error; a four-hour agent session that fails halfway is an incident. This is why cybersecurity review belongs in the design phase, not after the pilot.
Skills, Plugins and the Managed Agents Console
The other two nouns in the leak point at a marketplace, not just a runtime.
Skills as reusable capability
Skills are packaged instructions and assets an agent loads for a particular kind of task. Anthropic attaches them at the agent level, alongside the model, system prompt, tools and MCP servers. A skills layer is what turns Managed Agents from raw infrastructure into agentic AI a non-engineer can configure without touching code.
Plugins and the connector question
OpenAI already has the connector plumbing. AgentKit shipped a Connector Registry as a central place for admins to manage how data and tools connect across OpenAI products, gated behind the Global Admin Console. A plugin system for Managed Agents would most plausibly build on that rather than start again.
Managing many agents, not building one
The verb TestingCatalog highlights is “managing them.” That is a fleet-management problem: which agents exist, which environment each runs in, what each may touch, and who approved it. Anthropic’s answer includes scheduled deployments on a cron schedule and multiagent orchestration across threads.
The reference implementation is public
Anyone wanting to prototype against the likely shape of OpenAI’s product can read Anthropic’s documentation today. The concepts, event types and worker CLI flags are published, which makes the architecture unusually easy to anticipate.
The Ads Angle: When a ChatGPT Ad Resolves to Managed Agents
This is the part of the report with the largest commercial implications, and the weakest evidence.
What is reported
TestingCatalog says OpenAI is working on functionality that would allow ads in ChatGPT to resolve directly to an agent. Rather than an informational ad inside a response, the user would get a link to an agent that “takes the lead and guides them through the process.”
Why the timing is plausible
OpenAI announced on 31 August 2026 that ChatGPT Ads reached a $1 billion annualised revenue run rate in less than 200 days after launch, with tens of thousands of advertisers and availability in over 40 countries. Self-serve buying through Ads Manager was extended that day to India, Europe, the Middle East and North Africa.
The scale underneath it
OpenAI describes the ad-supported free tier as helping keep ChatGPT available to more than 1 billion weekly active users. An agent that completes a purchase, rather than an ad that describes one, is a different product category sitting on that base.
The competitive read
The report’s own framing is that agents handling conversion funnels could “pose a serious threat to major advertisers such as Meta and Google.” That is speculation and labelled as such, but the direction is consistent with OpenAI’s stated position that ads are one pillar of a diversified business alongside subscriptions, enterprise and usage-based APIs.
Price Is the Real Question for Managed Agents
Every other question about this product is downstream of what it costs.
The honest assessment of the incumbent
TestingCatalog is blunt that Anthropic “has led the pack in this area, although it remains difficult to say whether its managed agents have been super successful.” Developers and enterprises use them, but many users still choose cheaper alternatives. That is a demand signal, not a dismissal.
Where OpenAI has an advantage
The report identifies the obvious one: OpenAI already provides advanced models with strong computer-use capabilities, which is precisely what a sandboxed runtime consumes. Whether that converts into a price point that reaches critical mass is explicitly left open.
The precedent worth noting
When OpenAI shipped AgentKit, ChatKit, Evals and Connector Registry, all of them were included with standard API model pricing rather than sold separately. If Managed Agents follows that pattern and charges only for tokens and compute, it undercuts the category by default.
What to listen for on stage
The number that matters is not the token price. It is what a sandbox hour costs, whether idle sessions bill, and whether a self-hosted environment removes the runtime fee entirely.
Managed Agents Versus Rolling Your Own Loop
For most teams this is a build-or-rent decision with a clear dividing line.
When the managed route wins
Long-running tasks, secure sandboxes with network control, stateful sessions and scheduled runs are all things that take months to build well and are boring to maintain. Anthropic’s documentation names exactly these as the workloads its product is for.
When you should still build
Fine-grained control over the loop, unusual tool semantics, strict Zero Data Retention requirements, or a HIPAA BAA obligation all push you back to direct model access. The stateful design that makes a managed runtime useful is the same thing that disqualifies it from those regimes.
The lock-in question nobody asks early
An agent definition, an environment configuration and a session history are all vendor-shaped objects. Three product retirements in three years is a reasonable basis for asking how a large language model vendor plans to support this one, and for keeping your prompts and tool definitions portable.
| Factor | Managed runtime | Own agent loop |
|---|---|---|
| Time to first working agent | Days | Weeks to months |
| Sandbox and tool execution | Provided | You build and secure it |
| Data residency control | Only via self-hosted environments | Full |
| Zero Data Retention eligibility | Not on Anthropic’s beta today | Available |
| Cost predictability | Unknown for OpenAI | Your own infrastructure bill |
| Switching cost | High, vendor-shaped objects | Low, portable code |
What Could Go Wrong With Managed Agents
A hosted runtime concentrates several risks that are easy to underestimate in a keynote.
Autonomy without a clear owner
An agent with bash access, file operations and web fetch is an unattended employee with production credentials. Anthropic’s answer is permission policies with explicit tool-confirmation events; OpenAI’s equivalent is unannounced. Assume nothing until it is documented.
Compliance exposure from statefulness
Session history, sandbox state and outputs stored server-side are the source of the retention limits above. If a vendor stores the transcript to make sessions resumable, that transcript is in scope for every audit you face.
Deprecation risk on a young platform
The 30 November 2026 wind-down of Agent Builder is a documented fact, not a rumour. Any team standardising on Managed Agents should read that date as a description of the platform’s velocity.
Ads and agents in one system
If ads can resolve to an agent, the same runtime serves both a paying developer’s workload and a commercial funnel. Those two use cases have very different incentives around what an agent recommends, and that boundary deserves scrutiny long before it ships.
How to Prepare Before 29 September
There is useful work to do in the three weeks between the leak and the keynote.
Inventory what you already run
List every agent, script and scheduled job you have built on the Agents SDK, Agent Builder or a homegrown loop. Anything on Agent Builder needs a migration plan before 30 November 2026 regardless of what is announced.
Write down your environment requirements
Decide now whether a vendor-managed sandbox is acceptable, or whether you need self-hosted execution. That single answer determines whether Managed Agents is a candidate for you at all, and it should not be decided in the excitement of a launch.
Watch the keynote, read the docs
The keynote at 10:00 a.m. Pacific will give you the narrative. The documentation published alongside it will give you the beta header, the rate limits and the pricing, which is where the actual decision lives.
Managed Agents FAQ
Is Managed Agents officially announced?
No. As of 7 September 2026 it is a codebase finding reported by TestingCatalog. The agents UI is not accessible and OpenAI has not published anything about it.
When is OpenAI DevDay 2026?
Tuesday 29 September 2026 at Fort Mason in San Francisco, with a free livestreamed keynote at 10:00 a.m. Pacific featuring Sam Altman.
How would this differ from Workspace Agents?
Workspace Agents is a no-code product for business users inside ChatGPT, launched in April 2026. Managed Agents, on the Anthropic template, is a developer runtime you address through an API with your own agent definitions and environments.
Does Anthropic’s version cost extra?
Anthropic has not published separate Managed Agents pricing in its overview documentation. Its beta is enabled by default for API accounts, with MCP tunnels and dreaming held back in a narrower research preview.
Should we wait for DevDay before choosing a platform?
If you are on Agent Builder, no: that retirement date is fixed and migration work starts now. If you are choosing a new runtime, three weeks is a cheap wait for a documented price.
References
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.