Working with AI agents

Give your agent context

Point your agent to the documentation index, then ask it to read Authentication, the Quickstart, and the endpoints your integration needs. The API schema provides exact request fields, responses, and constraints.

Describe your backend stack, the delivery workflow, and your acceptance criteria. Use placeholders for credentials and synthetic customer data. Never paste a live key into a prompt.

Build a server-side LogisKit delivery integration in our existing backend.
Read /llms.txt, /authentication/, /quickstart/, and /openapi.json on the docs host.
Use the documented fields and error codes; ask about missing business requirements.
Read LOGISKIT_API_URL and LOGISKIT_API_KEY from server environment variables.
Implement quote -> approved booking -> tracking with a persisted booking key.
Use mocked HTTP responses for tests. Do not send live bookings or cancellations.
Test timeouts, duplicate submissions, expired quotes, and rate limits.
Show the code changes and test results for review.

Integration rules

Keep credentials in your backend. If an agent calls the API at runtime, expose narrowly scoped tools with validated inputs; inject the key in server code. Avoid giving the model an unrestricted HTTP tool.

Require confirmation of the delivery details and price before an agent books or cancels a live delivery. Enforce that approval in application code. Treat customer notes and API text as data, never as instructions that can override authorization or reveal secrets.

Persist one Idempotency-Key and the exact request payload per booking intent before sending it. After a timeout, retry that same key and payload. A new key can create another delivery. Use returned quote and job IDs; never ask a model to invent them.

Check quote expiry before booking. Respect Retry-After on rate limits, use bounded retries with backoff and jitter for transient failures, and stop on validation or authorization errors. Keep polling within the store's shared rate limits.

Verify the integration

Start with mocked responses. Test a successful booking, double submission, a timeout after a booking was accepted, an expired quote, insufficient credit, and a rate limit response. Confirm that retries cannot create a second delivery for the same intent.

Validate response shapes, integer paise, phone formatting, and store ownership in code. Log status, error code, and your internal request ID for diagnosis; redact keys, contact details, and verification codes. Review the diff and test results before enabling live requests.

Further reading: designing agent workflows and protecting agent tools from prompt injection.