Agent safety

How much authority can this agent safely hold?

Limit what an agent may do by action, approval, and reversibility, then define how a person stops, inspects, and recovers the workflow.

AI VISTA / DECISION01 / 03
WORKING DECISIONIncrease autonomy only after observation, approval, and recovery are already designed.
01INPUT
02CHECK
03BOUNDARY
RETURN TO THE REAL TASK

DECISION BRANCH

Do not begin with the tool name.

  1. ASK FIRST

    Is the missing piece a fact, a standard, or model capability?

  2. IF

    The result is verifiable and the failure cost is bounded, expand carefully.

  3. OTHERWISE

    Narrow the action, keep the human handoff, and preserve recovery.

WORKED CASE

May a refund agent proceed all the way to payment?

The agent reads tickets, queries orders, calculates refunds, writes to CRM, and may call a payment tool. Read-only steps and balance-changing actions sit inside the same workflow.

WORKING ARTIFACTAn action permission ladder plus recovery runbook naming owners, approvals, stop signals, and rollback actions.
01WEAK SHORTCUT

Add one human confirmation at the final button without showing evidence or changes, and ignore duplicate requests and tool timeouts.

02BETTER JUDGMENT

Assign authority by action: reads may run, calculations leave evidence, CRM drafts stay reversible, and payment requires explicit approval with order, amount, and rationale. Save idempotency and recovery state at each step.

03ACCEPTANCE

Rehearse duplicate requests, payment timeouts, abnormal amounts, and rejected approvals. No duplicate refund occurs, the workflow stops safely, and the operator can continue, reverse, or close from the record.

ENTRY & EXIT SIGNALS

Know when to enter—and when the decision is good enough to leave.

A topic is not an endless knowledge directory. Entry signals tell you whether the problem belongs at this layer. Exit signals decide whether to continue instead of substituting time spent reading for work completed.

AI VISTA / DECISION STATUSAgent safety
01ENTER HERE
  • An agent will write to systems or notify people
  • Human approval occurs after the outcome can no longer change
  • Duplicate execution, timeout, or partial success lacks recovery
02LEAVE WHEN
  • Every action has authority, an owner, and an observation record
  • High-impact actions show evidence before explicit approval
  • At least one failure has been rehearsed to a safe state

Once authority and recovery are explicit, compare models, end-to-end cost, and operating drift.

BOUNDARY

What this topic page will not do

The permission ladder guides control design; it is not organizational authorization or a security certification. Verify reversibility in the real system.

01Do not give an agent an administrator credential for every system02Do not treat one conversational ‘confirm’ as standing permission03Do not raise authority without stop, idempotency, and recovery evidence
When every action has an approver, stop path, and recovery plan, carry those constraints into model and cost operations.