Which integration boundaries should be agreed before a pilot starts?
A pilot integration agreement can define adapter ownership, schema compatibility and reproducible run manifests, reducing ambiguity when generated packages or host simulators change.
Agree which side supplies the environment adapter and which interface is authoritative. A pilot can use JSON Schema for observations and action arguments, plus a transition contract describing permitted state changes. Record whether the buyer's host or the supplied package implements those transitions.
Create a run manifest with `package_version`, `schema_version`, `adapter_version`, `runtime_version`, `seed` and `initial_state_digest`. These fields identify the combination actually evaluated. A seed alone is insufficient when different adapters interpret the same action differently.
Define reset through a concrete interface such as `reset(run_manifest)`, with a returned digest that can be compared against the expected starting state. Specify ownership of schema migrations and how unsupported versions fail. Silent field dropping can turn an integration problem into an apparent agent reasoning error.
Ask for a small conformance fixture containing an initial state, an action and its expected observable result. Run it before agent evaluation and after integration changes. The fixture checks the adapter boundary without substituting for business-quality inspection. Agree how mismatches are reported, including the manifest and trace needed to reproduce them, without presuming support commitments or service levels.
Make the package-to-simulator boundary inspectable so failures can be attributed to content, adapters or runtime behaviour.
- Audience:
- Environment builders
- Content type:
- Integration
- Buyer question:
- Scope and getting started
Content release: r-a220eb6e099b243b
Related questions
- Which licence questions should agent developers resolve before using a generated release?
- What can a focused pilot reveal about handling a challenged review finding?
- How should a buyer choose the scope of a first review pilot?
- What should a sample request specify so a buyer can judge fit without overextending the pilot?
- What technical handover should be agreed before connecting a generated release to an agent?
- What licence questions matter when generated assets are embedded in an environment?