Define the goal
Turn the request into concrete constraints, dependencies, and a structured task list.
Clear roles.
Observable work.
An agent workflow combines model-based analysis with constrained executable tools. This is a conceptual guide to responsibilities and evidence, rather than a live deployment status page.
Prepare the plan
Propose service configuration and request approved deployment tools.
Review the boundary
Inspect the origin policy, TLS settings, and scope of the proposed changes.
Check the evidence
Run checks, interpret observations, and report failures with a proposed next step.
Communicate in data.
Agents can exchange JSON messages or artifacts with a task identifier, sender, recipient, observations, and outcome. A downstream agent reads the previous output programmatically.
Explicit schemas let the application reject incomplete decisions before they reach a tool.
Separate reasoning from execution.
The model proposes or interprets an action. Deterministic code validates its arguments, checks approval, runs an allowed operation, and returns the actual result.
Failures become feedback. A tool error should produce a diagnosis or stop condition, not a fabricated success message.
Evidence matters. A useful execution log connects the task, model response, tool arguments, actual tool result, and validation outcome. Never include credentials in those records.