Your agent works. It’s been on a platform for months, the team uses it, and nobody has thought about the day it has to move. Then the vendor announces it’s switching off a component, or raises the price, or simply stops fitting. And you find out what was inside the platform that you thought was yours.
The thesis in one line: an agent’s lock-in isn’t in the model, which you can swap in an afternoon, but in what the platform keeps for you—flows, memory, evaluations, traces. If you don’t negotiate the exit on day one, you negotiate it on the day you have the least leverage.
How to leave an AI agent platform: what the platform holds that isn’t your prompt
When people talk about portability they talk about the prompt. It’s the most visible thing and the easiest to copy. But an agent in production is a system with parts that don’t fit in a text document, and most of them live on the vendor’s side. The first three—the prompt, the conversation history and the integrations—we already broke down in the guide to platforms for building AI agents, using what each vendor documents. Here we go after what that guide doesn’t cover: the assets that get generated afterwards, through use.
Which of your agent’s assets stay on the platform?
- Memory. What the agent has learned about your customers, your exceptions and your decisions. If it lives in a proprietary store, at best it comes out as text files and at worst it doesn’t come out at all. It’s the hardest asset to rebuild, because it was built by use. How to design it and where it should live: an AI agent’s memory.
- Evaluations. The set of cases you use to check the agent still does its job well, and the criteria that score each answer. Without them, migrating is blind: you can’t tell whether the new agent is as good as the old one. Measuring this continuously is a function, not a script: evaluating the quality of your AI agents.
- Traces. The record of what the agent did, with what data and why. It’s what answers an audit, an angry customer or your own team when something goes wrong. A platform that won’t let you export them takes away your evidence: traceability of AI decisions.
- Flows and logic. The agent’s definition as a graph of steps, tools and approvals. Sometimes it exports as code; sometimes only as a screenshot of a canvas. Those are two entirely different situations.
- Configuration and credentials. Permissions, connections, policies. An exported flow usually arrives without them and you have to recreate them one by one.
What happened this season: a platform switching off components
The case with a date is OpenAI’s. According to the vendor’s own documentation, Agent Builder—the visual agent canvas launched in October 2025—shuts down on November 30, 2026. Two June 2026 analyses (mcp.directory and Developers Digest) put OpenAI’s evaluations platform on the same schedule, going read-only on October 31; I couldn’t cross-check that with the official deprecations page, so read it before acting. What matters for this thesis is something else: neither analysis identifies a concrete mechanism for exporting evaluation datasets and criteria. A vendor can recommend that you migrate to the code SDK; it doesn’t owe you—unless it signs it—a tool to take away what you’d already measured.
Observability platforms don’t make it easy by default either. LangSmith’s documentation describes bulk export of traces to S3-compatible storage in Parquet format, with feedback comments as an opt-in field; the same page requires an Enterprise plan for accounts created after August 3, 2026. Exporting datasets as such isn’t listed on that page. It’s an example, not an accusation: portability exists, but it comes with a plan, a format and limits you need to know before you need it.
What does the law say if you’re a customer in the EU?
The Data Act has applied since September 12, 2025 and regulates switching between data processing service providers, a category that, according to the law firms that have analyzed it, includes software as a service. According to Faegre Drinker’s analysis (August 2025) and McCann FitzGerald’s (September 2025), the customer can switch providers with at most two months’ notice; the provider must complete the switch within 30 calendar days, extendable only if technically unfeasible; and switching charges are limited to direct costs until January 12, 2027, after which they can’t be charged. The provider must also allow “exportable data” and “digital assets” to be extracted in a machine-readable format, with one important exception: anything protected by the provider’s intellectual property or trade secrets falls outside.
Read it with two caveats. First: which of your assets counts as exportable data or as a provider’s secret is precisely what gets argued over, and there is no case law on agents yet. Second: this is information, not legal advice; your lawyer confirms what applies to you. But as a negotiating lever it works: it gives you the right to demand that the contract say what leaves and how.
Sources (accessed October 17, 2026): Faegre Drinker, McCann FitzGerald, OpenAI Agent Builder, mcp.directory, Developers Digest, LangSmith data export.
The six contract lines you negotiate on day one
- A closed list of what comes out. Flows, memory, evaluations, traces and configuration, named one by one and in an open format. “Your data” is not a list.
- A notice period for shutdowns. If the vendor retires a component, how many months’ notice and with what export window? Agent Builder was announced in June and shuts down in November: five months.
- A retrieval period and an erasure guarantee. How long you can download your data after termination and when it is destroyed. Get the confirmation in writing.
- Exit assistance at no extra cost. Who helps you empty out and under what conditions; in the EU, the 2027 cap on switching charges is your argument.
- Who owns what’s derived. Evaluations and memory were generated with your data and your judgment: have the contract say they’re yours.
- An export test in the first month. It isn’t a clause, it’s a habit: download everything in month one, open it and check that another system could read it. An export you’ve never opened doesn’t exist.
That last line is the one almost nobody does, and the one that makes the other five verifiable. It also belongs inside your governance and control of AI automation: someone, by name, is responsible for keeping the exit possible.
When don’t you need this much caution?
If your agent is a few-week experiment, lives in an assistant you already pay for and keeps no memory or evaluations of its own, the cost of leaving is low and negotiating six clauses is disproportionate: export the prompt and move on. Nor does it make sense to pay for a code framework to avoid a risk you haven’t measured yet. The rule is proportional: the more the agent learns from your business, the more it costs to have what it learned live in someone else’s box.
And if you doubt who will carry out the exit when it comes, that’s an operations question before a contract one: who maintains AI agents is where that answer stops being a name and becomes a function.
The line to take away: don’t ask which platform is best, ask what you’d take with you tomorrow if you left. If the answer is “the prompt”, you already know what you’ve left inside.