What integration contract should be agreed before connecting an agent to a pilot environment?

Before connecting an agent, agree the adapter contract between the buyer's runtime and the generated environment. Define tool names, argument types, required fields, error categories and whether each action reads or changes state. Specify how authentication roles are represented within the simulation.

Keep agent observations separate from evaluator access. The runtime might read account summaries and submit intake drafts, while the evaluator can inspect hidden acceptance conditions. Declare which fields are deliberately withheld so adapter code does not accidentally expose the answer.

The contract can identify schema_version, tool_contract_version and fixture_version independently. This lets buyers distinguish a changed business case from an interface change when comparing runs. Ask for migration notes where a revision changes field meaning or permitted actions.

For a first connection, request a minimal fixture and a documented reset response containing the active versions. Run a simple read-and-draft sequence, then inspect the stored result through the evaluator interface. This verifies the adapter before introducing complex cases. MirrorFirm could scope compatibility with the buyer's chosen runtime without implying that every framework or deployment configuration is already supported.