AI Agent Portability: What to Export So You Don't Get Locked In

AI Agent Portability: What to Export So You Don't Get Locked In

AI agent portability between vendors: export prompts, logs, memory and integrations so you can switch providers without losing control or starting over.

•

Want to apply this in your company?

Book a call with Juan →

AI agent portability is the ability to move an agent's operation to another vendor, platform, or model without starting from scratch. In practice, prompts, business rules, decision records, memory, integrations, and permissions must be stored in formats your company controls, not only inside a third party's dashboard. Before any switch, the safe path is to inventory these assets, export them, and test the new agent with real situations.

Key takeaways

  • The language model is the easiest part to replace. The operational value lives in the instructions, rules, history, and integrations around it.
  • Prompts and policies should leave your hands with a version, a review date, and the examples that explain why each rule exists.
  • Conversation and decision logs are company records, with a defined purpose and retention period, not just technical files kept by the vendor.
  • Memory and knowledge bases should come back as source documents, because indexes and embeddings are usually specific to each platform.
  • Credentials do not transfer. Each integration must be reconnected with minimum permissions in the new environment and tested before cutover.

Why vendor dependence is a problem today

The debate about superintelligence tends to dominate the headlines, but the most concrete risk for a company running agents is operational, and it shows up much sooner. Many teams start with a ready-made platform, build flows through the interface, and only later realize that the knowledge of the process is concentrated there. The problem is rarely the contract. It is the lack of a copy of what the agent does and how it was configured. When the vendor changes terms, discontinues a feature, or stops handling your use case well, the company has to rebuild everything from the memory of whoever did the setup.

Portability is therefore an architecture decision made early. It does not remove dependence on a model, but it lowers the cost of leaving. It also helps you compare vendors with discipline, because you can run the same set of situations on each one.

What to export: prompts, rules, and configuration

Start with whatever defines the agent's behavior. This includes:

  • General instructions and instructions for each step of the flow, in plain text, kept in a controlled repository or document.
  • Escalation rules: when the agent must stop, who receives the case, and in what format the handoff happens.
  • A list of allowed and forbidden actions, such as checking an order, changing an appointment, granting a discount, or promising a delivery date.
  • Response templates for sensitive situations, such as complaints, cancellations, and requests for personal data.
  • Parameters that do not appear in the prompt, such as response length limits, the order in which tools are used, and model settings.

Export a collection of test cases as well. It can be a spreadsheet with the customer message, the expected result, and the action the agent should take. This set is worth more than the prompt, because it shows the behavior the company has accepted and lets you check whether the new vendor reproduces it. To build these tests without exposing real customers, follow the approach in how to test AI agents with mocks and a safe sandbox.

Logs and decision history

Logs are the most underestimated part of portability. They record what the agent received, which tool it used, what response it sent, when it passed the case to a person, and what that person did next. Without this history, the company loses the ability to audit past decisions and to explain to a customer how an interaction happened.

When exporting, separate three kinds of records: the conversation transcript, the technical record of calls and errors, and the decision record, such as approvals, refusals, and handoffs. For the technical record, use a standardized structure. The OpenTelemetry semantic conventions for generative AI help name fields in a common way, which makes it easier to move the analysis to another tool.

Personal data in the history must not be exported without criteria. Names, phone numbers, and documents need a purpose, a retention period, and access control. If the switch involves moving data to another country or another processor, the international transfer analysis belongs in the plan, as explained in the guide on international data transfer under LGPD when using AI providers.

Memory and knowledge base

There is a common confusion between agent memory and knowledge base. The knowledge base is the company's documents: catalog, policies, lead time tables, and frequently asked questions. It should be kept as original files, outside the vendor's index. The index is a technical artifact that can be rebuilt. Embeddings generated by one model generally do not work in another without re-indexing, so what must be portable is the source text and the way it was split and labeled. Anthropic has described context engineering as curating what enters an agent's context window, which reinforces the value of keeping the origin of each piece of information.

User memory, such as preferences recorded for a customer, requires more care. Before exporting, confirm that the record has a legal basis and that the data subject can request correction or deletion. When the agent makes automated decisions based on that data, the right to review must also be respected in the new environment.

Integrations, credentials, and permissions

Integrations are where portability usually stalls. Each connection to a spreadsheet, CRM, calendar, or order system has credentials, permissions, and call rules that belong to the previous environment. The inventory must list each tool, its purpose, the type of access (read or write), who authorized it, and where the credential is stored.

When the integration follows an open standard, the switch becomes simpler. The Model Context Protocol, for example, describes a standard way to expose tools and data sources to language models. This does not eliminate the work of reconnecting permissions, but it reduces the rework of rewriting every connection. If your team uses Claude Code, the guide on configuring MCP in Claude Code helps you choose the right scope for each connection.

Permissions should be reissued, not copied. Apply the principle of least privilege in the new environment, and keep human approval for actions that move money, change customer records, or promise a deadline to a customer.

A switch plan in five steps

A switch plan can follow this order, which matches the method used in our projects: map, limit, pilot, and transfer.

  • Map: list the current flow, prompts, tools, log types, and knowledge base, with one owner for each item.
  • Limit: separate personal data, technical secrets, and what can be exported freely. Define which actions the agent may perform in the new environment before turning it on.
  • Export and compare: generate copies of the artifacts, run the test set on the current vendor and on the new one, and compare responses and actions taken.
  • Pilot: place the new agent in a real but limited flow, with full logging and a success metric agreed in advance, such as the share of handoffs that reach the right person.
  • Transfer or stay: if the metric is met, shut down the old agent on a set date. If not, adjust the configuration or keep the current setup, and record the reason.

Illustrative scenario: a building materials store uses an agent on WhatsApp to check stock and schedule deliveries. When evaluating another vendor, the team realizes the discount rules existed only in the current platform's interface. Before comparing any costs, they export the rules, create test cases with discount requests above the allowed limit, and confirm the new agent routes those cases to human approval.

How to decide between switching, adjusting, or staying

Before migrating, use a simple test for each situation:

  • Stay with the current vendor if the inventory is complete, the logs can be exported, and the test cases pass without major changes.
  • Adjust the configuration if the problem lies in the rules, the wording, or the limits, not in the platform.
  • Switch if the vendor blocks export of prompts, logs, or the knowledge base, or does not allow the decision record your operation requires.
  • Pause and review the project if the process is not yet mapped. An agent without a defined process has nothing to port.

This decision should be made by the person accountable for the process, not only by the technical team, because it involves risk, the cost of waiting, and responsibility toward customers.

Frequently asked questions

Can I change the language model without changing the agent platform? In many cases, yes, as long as prompts, rules, and test cases are stored outside the vendor. A model change still requires a new round of testing, because responses, formatting, and tool use can shift even with the same instruction text.

Do customer conversation logs have to be handed over to the new vendor? Not always. What the company should carry is the record it needs to keep, such as the decision and handoff history, with controlled access and a defined retention period. Before copying any conversation, verify the purpose and legal basis of the processing.

Do I need to rebuild the agent's memory from scratch? No, if the knowledge base was kept as original documents. Customer preference memory should be treated as personal data, and its export must respect correction and deletion requests. In most migrations, the main work is re-indexing the knowledge base in the new environment and validating the answers.

How do I know if switching vendors is worth it? It is worth it when the switch solves a measurable problem, such as blocked export, missing decision records, or insufficient quality on real cases. A limited pilot, with a metric agreed in advance and full logs, shows that difference with less risk than a full cutover on a fixed date.

Sources

AI agentsvendor lock-indata portabilityAI governanceoperations