Resource
Agent Workflow Risk Map
A practical review guide for agentic AI workflows. Use it to identify tool risk, context risk, recovery risk, and the human approval points required before production release.
This diagram shows the path from user goal to context risk, tool risk, approval, execution, recovery, audit tracing, and evaluation.
Tool risk
- Which tools can update records or trigger external workflows?
- Do high-impact tools require explicit human approval before execution?
- Are tool inputs validated against schemas before the agent can act?
Context risk
- Can the agent distinguish current context from stale, partial, or conflicting context?
- Does the workflow handle missing user intent without inventing assumptions?
- Are retrieval results, memory, and user-provided data separated clearly in traces?
Recovery risk
- What happens when a tool fails, times out, or returns an unexpected result?
- Can the user review, correct, retry, or cancel an agent action?
- Are failures logged with enough detail to debug the workflow later?
Review rule
An agent is not production-ready because it can call tools. It becomes production-ready when every risky action has a boundary, a trace, a recovery path, and a human approval decision where needed.