Documentation
Integration patterns for governed AI agents
A governed agent integration is a versioned contract that defines permitted operations, identity, data scope, authorization, retries, receipts, evidence and failure behavior.
An integration is not established by placing a vendor logo on a page. It is established by a supported operation, authenticated identity, accepted data scope, tested failure behavior and owned compatibility boundary.
This hub describes how integrations are scoped. Specific vendor and product availability is confirmed for the applicable release.
What integration families does an agentic platform need?
| Family | Purpose | Governance focus |
|---|---|---|
| Identity | Authenticate people, services and administrators | Delegation, revocation, strong privileged access |
| Data source | Retrieve records or documents | Tenant, classification, purpose, freshness and minimization |
| Model endpoint | Run inference | Approved provider, data terms, location, evaluation and fallback |
| Tool target | Create an external effect | Typed operation, approval, idempotency and receipt |
| Event source | Trigger work from changes | Authenticity, replay, ordering and duplicate handling |
| Evidence sink | Retain audit and receipts | Integrity, access, retention and export |
| Notification | Reach reviewers and operators | Recipient scope, content sensitivity and delivery receipt |
| Administration | Configure policies and adapters | Separation of duties and change evidence |
One provider may appear in several families. Each operation still needs its own contract.
What belongs in an operation contract?
operation: crm.contact.update
target_type: contact
effect: durable_write
identity: delegated_service_account
required_attributes: tenant, contact_owner, purpose
approval: resource_owner when material fields change
idempotency: request_id
receipt: target record version
timeout_state: unknown until reconciled
audit: request digest, policy, approval, receipt
The contract makes policy possible. A generic integration method such as call_api hides the consequence from the runtime.
How should identity be handled?
Prefer delegated or scoped credentials that identify the represented principal and permitted operations. Avoid one broad service account shared by unrelated workflows.
Document token lifetime, refresh, revocation, secret custody, administrator access, tenant scope and unavailable behavior. For disconnected deployments, identity and revocation sources must operate inside the boundary.
How should data access be handled?
Shape the query around tenant, ownership, classification and purpose before retrieval. Request only the fields needed for the task. Record source, freshness and data lineage.
If the integration returns more data than policy permits, filtering in the model prompt is too late. Enforce the boundary in the adapter or data store.
How should write operations be handled?
Use typed parameters, validation, current-state checks, parameter-bound approval, idempotency where the target supports it, and durable receipts.
For irreversible or externally visible actions, verify the final state from the authoritative target. A successful HTTP response may confirm receipt without confirming the intended business outcome.
What failure states must be defined?
| State | Required behavior |
|---|---|
| Authentication denied | Stop and alert the credential owner |
| Validation rejected | Return a specific recoverable error |
| Rate limited | Retry within a bounded policy or hold |
| Timeout before send | Safe retry may be possible |
| Timeout after send | Reconcile before retrying |
| Partial batch | Record per-item state and continue only by policy |
| Duplicate event | Detect and suppress or handle idempotently |
| Target version changed | Re-read and require a new decision when material |
| Evidence write failed | Hold consequential action or enter explicit degraded policy |
| Unknown response | Preserve ambiguity and assign reconciliation |
Calling every non-200 response "failed" can cause a duplicate side effect when the target accepted the request but the response was lost.
How are integrations tested?
Test authentication, least privilege, positive operations, prohibited operations, malformed parameters, cross-tenant access, expired grants, duplicate delivery, rate limits, timeouts, partial success, API-version change, unavailable target and evidence failure.
Run tests against a representative sandbox or faithful contract fixture. Record the adapter and API versions. A prior green test does not establish current compatibility after either side changes.
How should integration availability be published?
Publish specific support only when the release evidence and commercial offer agree. State whether an integration is built-in, configured, custom, preview, read-only or planned. Name the supported operations and product.
This company page deliberately does not infer support from providers mentioned elsewhere in portfolio copy. Contact Trunnion with the target system and operations to receive a current scope.
What should a buyer provide for scoping?
Provide the product and API documentation, authentication method, desired operations, object and field scope, approximate volume, data classification, environment, test account, error and retry expectations, acceptance criteria, and the owner who can approve target-system changes.
Do not send secrets or restricted data through the public site. Establish an authorized channel before exchanging credentials, exports or production examples.
Direct answers
Frequently asked questions
Which vendors does Trunnion integrate with?
Integration availability is product and release-specific. This company page documents integration patterns rather than publishing a logo list that could imply unverified support.
What makes an integration governed?
The runtime uses typed operations, least-privilege credentials, contextual authorization, approval where required, idempotency or reconciliation, and linked execution evidence.
Can Trunnion connect to a custom API?
A custom integration can be scoped when the API, authentication, permitted operations, data classes, environments, test system and ownership are available.
How are integration failures handled?
Each operation defines timeout, retry, duplicate, partial-success and unknown-outcome behavior. Consequential actions require reconciliation with the authoritative target.
Are read and write integrations treated differently?
Yes. Writes, transmissions and external commitments generally require narrower authority, stronger approval and durable receipts than low-sensitivity reads.
How are integration claims verified?
The evidence should name the product release, adapter version, target API version, supported operations, tested environment, negative cases and known limitations.
