Boundary

Give the agent the smallest useful action space

Use scoped credentials, narrow tool access, spending and rate limits, and explicit prohibited actions. Do not give broad access “for convenience”.

Open responsible source ↗

Accountability

Keep meaningful human control

Place approval before external publication, payment, deletion, sensitive communication or other consequential actions. Make the approver’s evidence clear.

Open responsible source ↗

Operations

Observe, interrupt and recover

Log decisions and tool calls, detect unusual behaviour, provide a kill switch and design compensating actions for failures.

Open responsible source ↗

People

Tell users when an agent is acting

Use plain disclosure, train operators and provide a route to question or challenge an outcome.

Open responsible source ↗

FREQUENTLY ASKED

Two points worth making clear

What makes an AI system “agentic”?+

It can pursue a goal through multiple steps and use tools or systems with some autonomy, rather than only returning one answer.

Is a human-in-the-loop enough?+

Not by itself. The person needs timely control, understandable evidence and real authority to stop or reverse the action.

NEXT STEP

Apply the framework to your actual task

Desk AI can help structure evaluation questions and controls. It does not make procurement decisions or retain chat data.