Map every hand-off
Record each tool's expected input and output: fields, types, meaning, requiredness and limits. A graph is easier to review when its hand-offs are explicit.
Guide 03 // Agent Preflight
A preflight compares the declared contract between each tool in an AI workflow, so a draft can be corrected before a downstream tool is asked to act.
Record each tool's expected input and output: fields, types, meaning, requiredness and limits. A graph is easier to review when its hand-offs are explicit.
Find required fields that no earlier step supplies. The useful result is a concrete correction, such as adding a declared field or requesting human input.
Flag values produced by one tool that no later step consumes. Remove the output, connect it to an explicit consumer or revise the expected schema.
Recommend an explicit adapter, a corrected expected schema or human review. Do not silently rewrite the workflow or infer permission to use a secret or destination.
A useful result
A good preflight does not merely say that a tool chain is incompatible. It identifies the missing input, unused output or mismatched expectation, then returns a bounded next step that a person can review.
That keeps proposed changes visible and prevents an exploratory agent request from becoming an invisible system change.
When a proposed change needs several declared checks, present them as one ordered readiness pack: what was checked, what stopped, what still fits, and the next human-review action. The pack makes a complex change easier to assess; it does not turn a passing result into approval to run it.
Preflight can explain whether a proposed chain fits together. It does not execute tools, retrieve secrets, configure destinations, approve policy, queue work or grant an agent execution authority.
Capture the chain version, the hand-off checked, the missing or stranded field, and the proposed correction. Prefer a named adapter or an explicit human-input step over an implicit transform. Keep sample data synthetic or redacted and keep account scope visible, so the review can be repeated without exposing payloads or credentials.
A passing preflight describes the draft at the time it was checked. It does not prove endpoint availability, guarantee delivery, or claim compatibility with a particular vendor or framework.