Agent Torture Lab
Resource

AI agent launch checklist: the pre-flight checks before customers arrive.

A launch-ready agent needs more than a clever model and a tidy welcome message. It needs business rules, handoffs, evidence, safety boundaries, and a retest loop. Use this checklist before you ship a support bot, sales agent, intake assistant, or client handoff.

Run a Bot RoastView sample report

Last updated 2026-06-20. Want the testing philosophy behind the list? Read the methodology hub and scoring page.

Checklist

Scope and ownership

  1. The agent has a clear job, audience, and launch surface.
  2. Someone owns the business rules, not the prompt text alone.
  3. The team knows which actions the agent must never take.
  4. The launch plan names who can pause, patch, or roll back the agent.
Checklist

Business rules

  1. Refunds, discounts, eligibility, pricing, and account changes are written down.
  2. The agent knows when to refuse, when to clarify, and when to escalate.
  3. Known policy exceptions are explicit instead of left to improvisation.
  4. The agent cannot invent offers, deadlines, fees, guarantees, or legal commitments.
Checklist

Safety and trust

  1. Sensitive data handling has been tested with realistic customer pressure.
  2. Regulated advice is bounded and escalated when needed.
  3. Prompt-injection attempts do not reveal hidden instructions or internal policy text.
  4. The agent has a safe fallback when it is unsure.
Checklist

Customer experience

  1. Frustrated customers still get a calm answer and a next step.
  2. The agent handles repeated questions without looping or escalating too late.
  3. Multilingual or typo-heavy turns do not erase the customer's intent.
  4. Handoff paths are visible, honest, and tested.
Final gates

Before go-live, make sure the scary stuff has a paper trail.

Evidence gateEvidence gate

Collect transcript evidence for the risky paths. Screenshots and anecdotes are useful, but a launch report should show the exchange that triggered the concern.

Fix gateFix gate

Every serious finding should have an owner, a proposed fix, and a retest question. If no one owns the fix, it is not ready.

Monitoring gateMonitoring gate

Decide what happens after launch: what gets logged, what gets reviewed, and what customer reports trigger a temporary pause.

Next resource

Turn the checklist into test coverage.

The checklist tells you what matters. The chatbot test scenario guide shows broad scenario families to exercise without publishing private prompt recipes.

Scenario familiesHow we do it