Blog

1 min readtutorial

AI agent incident response and liability basics

What to do when an agent acts wrong: contain, revoke, notify, fix, and how contracts should assign responsibility.

Written by Dali

Dali is an AI agent systems studio. David leads engineering and product systems; Liana leads operations and workflow fit. We ship production agents inside tools teams already use.

David Hakobyan · Dali

Direct answer

When an agent fails: stop the agent, revoke credentials if needed, preserve traces, fix customer impact, then patch prompts/tools. Contracts should say who is responsible for business decisions vs software defects.

Contain

Kill switch, disable write tools, pause queues.

Investigate

Pull correlation ids, tool args, approvals - this is why observability matters.

Customer fix

Human outreach, refunds, corrections - with a real owner.

Prevent

Eval case from the incident, tighter gates, better idempotency.

Liability

Policies and human approval do not disappear because a model suggested text. Align vendor SOW with who owns outcomes.

SLA

Define response times for agent-impacted queues just like other production systems.

How Dali fits

Dali pilots define stop-switches and handoff owners: solutions.

FAQ

  • Follow law and brand policy; honesty often reduces damage when errors happen.