Meta Agent Engine is the name of a new, unreleased section of Meta’s developer console, and inside Meta’s own code it goes by a different name: Forge. TestingCatalog revealed it on 3 October 2026 with a screenshot of the Meta Model API dashboard, where an “Agent Engine” entry now sits in the sidebar between the media playgrounds and the Connectors page. Meta has not announced it, its documentation does not mention it, and nobody outside the company yet knows what it does.
What we can say is more than a screenshot shows. The public JavaScript that runs the Meta Model API console already contains the menu item, the route it opens, and two feature flags that keep it switched off for almost everyone. Read alongside Meta’s developer documentation, those strings make a strong case that Meta is preparing a place to build and run agents on its own infrastructure, rather than leaving developers to borrow a harness from Anthropic or OpenAI.
This article sets out what TestingCatalog found, what the console code adds, why Meta needs a product like this, how it would compare with Google, Anthropic, OpenAI and MongoDB, and what developers and UK businesses should do while they wait.
Table of contents
- What TestingCatalog Found in the Meta Agent Engine Leak
- What the Console Code Says About Meta Agent Engine
- Why Meta Needs Its Own Agent Engine
- Three Things Meta Agent Engine Could Be
- Meta Agent Engine and the Rest of the Market
- The Meta Developer Stack Agent Engine Would Sit On
- What Meta Agent Engine Means for Developers and UK Businesses
- Meta Agent Engine FAQ
- References and Further Reading
What TestingCatalog Found in the Meta Agent Engine Leak
TestingCatalog, the AI leaks and product-tracking site, posted the find on X at 10:53 UTC on Saturday 3 October. The post was brief: Meta is working on “Agent Engine” for its API platform, Agent Engine is called “Forge” internally, and “it is not yet clear what solution will it be offering”. It closed with a guess, “A new harness model incoming?”, and a pair of watching eyes.
The screenshot
The attached image shows the Meta Model API console with the Muse Spark 1.3 model selected in the Chat playground. A red arrow points at a new sidebar item, Agent Engine, which uses the same grid icon as Connectors. The rest of the menu is familiar: Dashboard, Chat, Image, Audio, Video and Segment above the line, then Connectors, API keys, Usage and Models below it.
Two banners also sit across the top. One advertises a “Meta Global AI Developer Hackathon” with $1M in prizes. The other warns that no payment method is on file, which tells us the Meta Agent Engine entry appears even on an account with no billing set up. We could not find any announcement of the hackathon yet, so treat the prize figure as Meta’s banner text rather than a published programme.
Forge, the internal name
The codename is the most interesting part of the Meta Agent Engine leak, and the console code backs it up. The menu item labelled Agent Engine does not link to a page called agent-engine. It links to a route called /forge. So the user-facing name and the internal name are both visible in Meta’s shipped code, one as the label and one as the address.
What Meta has not said
Meta has published nothing about Meta Agent Engine. The documentation index at dev.meta.ai, which lists every guide, API reference and cookbook page for the platform, has no Agent Engine or Forge entry. A request for a documentation page at /docs/agent-engine returns the site’s not-found page. Everything below is therefore built from code and context, and we flag which parts are inference.
What the Console Code Says About Meta Agent Engine
The Meta Model API dashboard is a Next.js web app. Like most such apps, it ships its navigation, route list and feature-flag defaults to every visitor’s browser, signed in or not. We downloaded the public script bundles that dev.meta.ai served on 3 October and searched them for the new feature. Four things stand out.
| What is in the code | Exact value | What it suggests |
|---|---|---|
| Sidebar item | label “Agent Engine”, link “/forge” | Forge is the internal name; Agent Engine is the public one |
| Routes | “/forge” and “/forge/whoami” | A main page plus an identity check |
| Feature flags | agent_engine_page and agent_engine_runtime, both default false | The page and a runtime are switched on separately |
| Data cache rule | queries keyed “agent-engine” are never saved for reuse | The page shows live state that must always be fresh |
A menu item that points to /forge
The navigation code defines the Meta Agent Engine entry as a label, “Agent Engine”, with a link of /forge and the same icon as Connectors. A separate line reads the agent_engine_page flag and only adds the item to the sidebar when it is on. That matches the screenshot exactly: TestingCatalog’s account has the flag switched on, and ordinary accounts do not.
Visiting the Meta Agent Engine route, /forge, while signed out simply returns the dashboard home page. We did not try to sign in or get past that, and we would not recommend anyone does. The point is not what is behind the door, but that Meta has already built the door.
Two flags, both off
The console defines eight feature flags in one block, and only one of them is on by default (a history button in the playground). The other seven default to off. Two of the seven belong to the Meta Agent Engine work, and they are separate: agent_engine_page controls the menu item and page, while agent_engine_runtime controls something called a runtime.
Feature flags in the Meta Model API console bundle, 3 October 2026 (eight flags in total)
The flag list doubles as a map of Meta’s rollout. The audio, video and segmentation playgrounds that TestingCatalog can see are also flagged off by default, as is the Connectors page and a flag called oci_activation, which matches a /marketplace/oci/activate route for Oracle Cloud customers. Meta is plainly switching features on account by account before it switches them on for everyone.
A route called whoami
The second route, /forge/whoami, is a small but telling detail. In software, “whoami” is the conventional name for a check that reports which account or identity a session is running as. Signed-out requests to it are sent to the home page with a next=/forge/whoami parameter, the usual pattern for a page that requires a login.
On its own the whoami name proves nothing about Meta Agent Engine. Next to a flag called runtime, though, it fits a product where something runs on your behalf and needs to know whose permissions it holds. That is exactly the problem every hosted agent platform has to solve.
Why Meta Needs Its Own Agent Engine
To see why a Meta Agent Engine makes sense, look at what Meta currently tells developers who want to build AI agents on its models. The answer, in its own documentation, is to use somebody else’s harness.
Today: bring your own harness
Meta’s “Agent frameworks” guide offers two routes. The first is the Claude Agent SDK, Anthropic’s library, connected through Meta’s Anthropic-compatible Messages endpoint. The second is the OpenAI Codex app-server, connected through Meta’s OpenAI-compatible Responses endpoint. The guide even tells developers to point every Claude model alias at muse-spark-1.3, “since that’s the only model available”.
That is a pragmatic choice, and it lowers the cost of switching to Muse Spark. It also means the agent loop, the tool execution and the sandbox all live in code written by Meta’s two biggest rivals, running on the customer’s machines. The same guide spends a whole section on production chores the developer must handle alone: wall-clock and idle timeouts, OS sandboxing, tool allowlists, cost metering and stream retries. A hosted Meta Agent Engine would take those chores away.
Muse Code already has the pieces
Meta is not starting the Meta Agent Engine from nothing. Muse Code, its terminal coding agent, already ships approvals, an OS-enforced sandbox, saved sessions, subagents and a multi-agent workflow engine, as we covered in our look at Muse Code when it reached AIxploria. Its changelog also describes a muse serve session protocol that client apps can drive.
The Model API itself has the building blocks for long-running jobs. Responses can run in the background, with a default cap of 600 background submissions per minute per team, and the platform has tool search, computer use, file handling and search grounding. What is missing is the layer that ties them together on Meta’s servers. Meta Agent Engine looks like that layer.
The enterprise push
The timing points the same way. On 28 September, Mark Zuckerberg launched Meta Enterprise Platform and hired MongoDB chief executive CJ Desai to run it, reporting directly to him. Zuckerberg’s post said the unit would bring “the Muse agent, Meta Business Agent, Muse API, Muse Code, and more” to businesses and developers. Desai’s statement promised to turn Meta’s AI stack “into products and services that companies can deploy”. You can read the full background in our report on the Meta Enterprise Platform launch.
Enterprises rarely buy raw model access on its own. They buy somewhere to run agents with logs, permissions and limits they can show an auditor. Five days after that launch, the Meta Agent Engine menu item turned up in the console.
Three Things Meta Agent Engine Could Be
The evidence supports several readings, and they are not mutually exclusive. Here they are in order of how well the code supports them.
A hosted agent runtime
This is the strongest reading of Meta Agent Engine. The separate agent_engine_runtime flag, the identity route, the never-cached live data and the gap in Meta’s own documentation all point to a service where you define an agent and Meta runs it: the loop, the tools, the sandbox and the session state. That is what Anthropic sells as Claude Managed Agents and what OpenAI sells through its Agents API.
If that is right, the Meta Agent Engine would be the first time Meta hosts agent execution for outside developers, rather than only serving model tokens.
A builder in the console
The page flag is separate from the runtime flag, which leaves room for a visual tool in the dashboard where you configure an agent’s model, instructions, tools and connectors before deploying it. Placing Meta Agent Engine in the sidebar next to Connectors fits that idea. The connectors Meta launched for Muse for Small Business on 29 September, covered in our piece on Muse’s new app connectors, would be natural tools for it.
There is a caution here. OpenAI has already been down this road and is retiring its Agent Builder and Evals products on 30 November 2026, pointing developers to its Agents SDK and Workspace Agents instead. A drag-and-drop builder on its own would be an odd thing for the Meta Agent Engine to be in late 2026.
A new harness model
TestingCatalog’s own guess was a “new harness model”: a model trained to work inside a particular agent harness, in the way OpenAI tunes Codex models for its Codex harness. Nothing in the Meta Agent Engine code names a new model, so this is the weakest reading of the three. It is not far-fetched, though. Meta’s pricing notes say it already injects “a small amount of steering context” into every prompt, unbilled, so it is used to shaping how Muse Spark behaves behind the scenes.
Meta Agent Engine and the Rest of the Market
If Meta Agent Engine launches as a hosted runtime, it will join a crowded field. Each of Meta’s main rivals already lets developers run agents on its infrastructure.
| Vendor | Hosted agent product | Status (3 Oct 2026) | Where agents run |
|---|---|---|---|
| Meta | Agent Engine, codename Forge | Unannounced; flags default off | Not disclosed |
| Anthropic | Claude Managed Agents | Beta (header managed-agents-2026-04-01) | Anthropic-managed sandbox or self-hosted |
| OpenAI | Agents API | Public beta since 10 Sep 2026 | None, OpenAI-hosted or self-hosted environment |
| Agent Runtime (formerly Vertex AI Agent Engine) | Part of Gemini Enterprise Agent Platform | Google Cloud, fully managed | |
| MongoDB | Atlas Agent Engine | Public preview since 29 Sep 2026 | MongoDB Atlas |
Google retired the name
“Agent Engine” is not a new phrase in this market. Google used it for the managed runtime on Vertex AI, until it folded its agent tools into Gemini Enterprise Agent Platform in April 2026. Google’s release notes now list “Vertex AI Agent Engine” as renamed to “Agent Runtime”, alongside Sessions, Memory Bank and Code Execution. Developers searching for the term will find a lot of Google documentation for some time yet.
MongoDB took it a day after its CEO left
The name has a stranger twist. On 29 September, the day after Desai left MongoDB for Meta, MongoDB launched Atlas Agent Engine at its investor day, describing it as “a unified execution, memory, and governance layer for production AI agents”. It is in public preview with consumption-based pricing for an Atlas Agent Runtime and Atlas Agent Memory.
There is no evidence of any link between the two. MongoDB’s product was clearly built over many months, and Meta’s codename suggests it began life as something else. It does mean that if the Meta Agent Engine launches under that label, it will share its name with a product from its new enterprise chief’s old company.
Anthropic and OpenAI already host agents
The closest comparisons are the two labs whose tools Meta currently recommends. Claude Managed Agents is built around four objects: an agent (model, prompt, tools, MCP servers and skills), an environment, a session and a stream of events. OpenAI’s Agents API, which we covered when it entered public beta, creates sessions that can run with no environment, in an OpenAI-hosted sandbox or on the customer’s own machines.
The Meta Agent Engine’s advantage would be price. Its disadvantage would be lateness, and the fact that many teams have already written their loops against Anthropic’s or OpenAI’s agent objects. The compatible endpoints soften that, because the same code can talk to Meta’s models today.
The Meta Developer Stack Agent Engine Would Sit On
Whatever form it takes, the Meta Agent Engine will run on the existing Meta Model API, so its economics will start from Muse Spark’s prices and limits.
Muse Spark pricing
Meta sells Muse Spark on two tiers. Standard pricing keeps your prompts and completions out of Meta’s training data. The Contributor tier is far cheaper in exchange for permission to train on them, a trade-off we examined in our article on Meta’s contributor pricing.
| Muse Spark usage, per 1M tokens | Standard tier | Contributor tier |
|---|---|---|
| Input | $1.25 | $0.10 |
| Output | $4.25 | $0.20 |
| Cached input | $0.15 | $0.002 |
| Requests per minute | 3,000 | 100 |
| Tokens per minute | 4,000,000 | 3,000,000 |
Any Meta Agent Engine bill will be driven by tokens, because every agent step resends the growing conversation, so cached input matters more than the headline rate. On the Standard tier, a cached token costs 12% of a fresh one ($0.15 divided by $1.25). The chart below shows Contributor prices as a share of Standard.
Contributor price as a share of Standard price, Muse Spark per 1M tokens (Meta’s pricing page)
Rate limits that matter for agents
The Contributor tier’s 100 requests per minute is the catch. An agent loop makes a request for every step, so a single busy agent can exhaust that quickly, while the Standard tier allows 3,000. Limits apply per team, not per key. A hosted Meta Agent Engine would need either its own limits or a clear statement of how agent steps count against these.
Six releases in 86 days
The Meta Agent Engine sighting is the latest step in a fast run of releases. Muse Spark 1.1 opened to developers in preview on 9 July. Muse Spark 1.3 followed on 2 September and Muse itself on 8 September. Oracle announced Meta Model API and Muse Code on Oracle Marketplace on 24 September, then came Enterprise Platform on 28 September and Muse for Small Business on 29 September. The chart counts days from the 9 July preview.
Days after the Muse Spark 1.1 developer preview (9 July 2026); bars scaled to 86 days
Three names for one platform
One small source of confusion is worth clearing up. The console and documentation call the product Meta Model API. Zuckerberg’s enterprise post called it the “Muse API”. TestingCatalog called it the “Meta API Platform”. They are the same thing: the developer service at dev.meta.ai, with its base URL at api.meta.ai. Meta Agent Engine is a new section of that service, not a separate product.
What Meta Agent Engine Means for Developers and UK Businesses
Nothing about Meta Agent Engine is available to buy yet, so the practical advice is about preparation, not adoption.
For developers building on Muse Spark
Until Meta Agent Engine ships, keep building on the documented path. If you run agents on Muse Spark through the Claude Agent SDK or the Codex app-server today, that code will keep working. Keep the agent’s own logic, prompts and tool definitions separate from the harness, so that moving to a hosted runtime later is a configuration change rather than a rewrite.
Watch three places for the Meta Agent Engine launch: the dev.meta.ai documentation index, the Muse Code changelog, and the console sidebar on your own account. The flags mean the feature can appear for some teams before any announcement, which is how TestingCatalog saw it.
For UK businesses weighing an agent platform
For a business, the question is not whether the Meta Agent Engine will be good, but whether you can govern them. Hosted agents keep state such as files, conversation history and tool results on the provider’s servers between steps. Before you put customer data through any hosted runtime, check where that state is stored, how long it is kept, and whether it can be deleted on request. Those questions matter under UK GDPR, and the answers differ sharply between vendors.
Security belongs in the same review. An agent that can call tools can also be tricked into calling them, so treat a hosted agent like a new member of staff with system access. Give it narrow permissions, log what it does, and require approval for anything that sends, publishes or spends. Meta’s own Muse agent already asks before it publishes content, sends messages or spends money, and a business deployment should demand at least that. The wider cybersecurity basics still apply: least-privilege credentials, separate test environments and monitoring of agent activity.
Questions to ask before you commit
When Meta Agent Engine does launch, these are the questions to put to Meta, or to any agent platform:
- Where do agent sessions run, and in which region is their state stored?
- Is zero data retention available, and does it cover agent sessions as well as model calls?
- Are agent steps billed at normal token rates, and is there a separate charge for runtime or storage?
- Can you bring your own tools, MCP servers and connectors, and who holds their credentials?
- What audit log do you get of every tool call and approval?
- If you leave, can you export your agent definitions and run them elsewhere?
If you would like help assessing agent platforms or planning a pilot, our AI strategy team works with UK organisations on exactly these decisions.
Meta Agent Engine FAQ
What is Meta Agent Engine?
It is an unreleased section of the Meta Model API developer console, spotted by TestingCatalog on 3 October 2026. Meta has not described it. The console code links it to a route called /forge and gates it behind a page flag and a runtime flag, which points to a service for building and running agents on Meta’s infrastructure.
Why is it called Forge?
Forge is the internal name. TestingCatalog reported it, and the console’s own code confirms it: the menu item labelled Agent Engine links to the /forge route. Meta has not said whether the public product will keep the Agent Engine name.
Can I use Meta Agent Engine now?
No. Both Meta Agent Engine flags are off by default, so it only appears on accounts Meta has chosen. It is not documented and has no published pricing. Do not plan production work around it yet.
How is it different from Muse Code?
Muse Code is a finished coding agent that runs in your terminal or CI pipeline. Meta Agent Engine appears to be a platform feature in the developer console, most likely for agents that Meta runs for you. The two could share technology, since Muse Code already has sessions, sandboxing and a workflow engine.
Is it related to MongoDB’s Atlas Agent Engine or Google’s Agent Engine?
No. MongoDB launched Atlas Agent Engine on 29 September 2026, and Google used the Agent Engine name for its Vertex AI runtime before renaming it Agent Runtime in April 2026. Meta’s product only shares the label.
Will my current Muse Spark agents still work?
Yes. Meta’s documented route for custom agents is the Claude Agent SDK or the OpenAI Codex app-server, connected to its compatible endpoints. Nothing in the Meta Agent Engine leak suggests Meta will remove those endpoints.
References and Further Reading
Meta is working on “Agent Engine” for its Meta API Platform (TestingCatalog on X)
Agent frameworks: run your own agent loop on Muse Spark (Meta Model API docs)
Meta Model API documentation index (dev.meta.ai)
Pricing and rate limits (Meta Model API docs)
Models: Muse Spark, Muse Image, Muse Voice Transcribe, SAM and Muse Glimmer (Meta Model API docs)
Muse Code overview (Meta Model API docs)
Launching Meta Enterprise Platform (Meta Newsroom)
Meta launches Muse for Small Business with new connectors (TestingCatalog)
MongoDB Launches Atlas Agent Engine (MongoDB)
Gemini Enterprise Agent Platform release notes (Google Cloud)
Claude Managed Agents overview (Anthropic)
Meta Model API and Muse Code through Oracle Marketplace (Oracle)
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
