What tool contract could support generated referral intake tasks?
A generated intake API can expose structured referral creation, field validation and recoverable errors, allowing labs to train tool use alongside customer communication.
A task pack could define a create_referral action accepting account_id, source_contact_id, reason_code, customer_summary, contact_preference and policy_version_id. Fields should have documented types, allowed values and null semantics. A missing contact preference should be distinguishable from a customer declining to provide one.
The tool response can return referral_id, status and validation_errors. A rejected submission might identify an unsupported reason_code or a required field that is absent. The model must use that response to repair the request or obtain the missing information, rather than describe the referral as completed.
Generated fixtures can pair request JSON with response JSON and the resulting record state. A schema file can define whether unknown fields are rejected, ignored or retained, so behaviour does not depend on an undocumented parser assumption.
An evaluation hook can capture submitted arguments, validation feedback and the next model action. This supports inspection of error recovery: whether the model corrects the specific defect, avoids inventing customer information and reports the actual submission result.
Make referral submission executable through a documented contract with inspectable inputs and outcomes.
- Audience:
- Model labs
- Content type:
- Integration
- Buyer question:
- What can be generated
Content release: r-a220eb6e099b243b
Related questions
- Which variations should a generated intake pack contain?
- What generated context can surround a customer's request for assistance?
- How can generated account state be packaged for repeatable agent runs?
- Can a generated pack include conflicting policy documents and outdated routing instructions?
- How can generated tasks connect premium difficulty with a separate insurance claim?
- Which generated records help simulate a credible referral without exposing every detail?