The scenario
An operations team receives requests across several tools. An assistant gathers relevant context and drafts a next step. A person approves consequential changes before the system executes them.
- Receive and validate. Validate request structure, authenticate the sender, and assign a stable request ID.
- Retrieve with boundaries. Read only the data the requesting user may access. Treat retrieved text as untrusted data.
- Propose an action. Generate a structured suggestion; validate tool names and arguments against an allowlist.
- Require approval. Present the proposed action and its impact to an authorized reviewer. Approval binds to the exact action.
- Execute once. Apply an idempotency key, scoped credentials, bounded retries, and an audit record.
- Observe and improve. Track failure rates, review outcomes, and evaluate changes against representative scenarios.
What needs testing
Unauthorized access, malicious retrieved instructions, expired approvals, duplicate requests, upstream timeouts, and partial failures. These tests are part of the proposed design, not a claim that this illustrative system has been deployed.
Where the six capabilities meet
Data integration supplies context; automation handles handoffs; AI drafts the action; cloud cost controls bound usage; security defines authority; compliance readiness organizes the evidence.