Agentic AI
Delegate choices, not accountability.
There is no single industry definition of an AI agent.
In this course, an agent is an AI-enabled system in which a model chooses at least some next steps or tool calls inside an authorized loop. A fixed sequence remains a workflow, even when some steps use AI.
Capability is not permission
Tools can let a system retrieve information or change external state. What an agent can technically do is different from what it is authorized to do.
- Suggest: Recommend an action for a person to perform.
- Prepare and pause: Draft or stage an action for approval.
- Act within bounds: Complete low-risk, reversible actions within explicit limits.
Use the least autonomy that completes the job.
A simpler prompt, deterministic tool, or fixed workflow is often easier to test and govern.
Define the boundaries before the loop begins.
- One bounded job and completion condition
- Minimum data and tools required
- Separate read and write access
- Limits on scope, time, retries, and spending
- Human approval points
- Representative evaluations and monitoring
- Isolation from untrusted external content
- A stop, revocation, and recovery plan
Accountability does not transfer to the tool.
Name who can approve, stop, recover, and own the result.
Write an agent charter.
Define the job, inputs, tools, forbidden actions, approvals, limits, evidence log, and recovery plan. If any item is unclear, return to Workflow Engineering.
Autonomy is a design variable. Accountability remains human.
Return to the four-level frameworkSource checks: Anthropic, Building effective agents; OpenAI, A practical guide to building agents; NIST AI RMF Core. OPEN QUESTION: Recheck the NIST citation before launch because AI RMF 1.0 is under revision.