> ## Documentation Index > Fetch the complete documentation index at: https://docs.armin.cx/llms.txt > Use this file to discover all available pages before exploring further. # Testing and run history > Test AI Agents before launch and inspect what happened after each run. Tests show how an AI Agent will handle realistic support cases. Run history shows what it did after a test or live trigger. [Open run history](https://armin.cx/app/_/ai/agents-v2/runs) to inspect executions across all agents. ## What to test Test the agent with messages that represent real customer situations: * the happy path the agent should handle * messages that should escalate * ambiguous customer requests * missing order or customer data * sensitive actions that require approval * messages that should not trigger the agent ## Test ownership and hand-offs Use a short ticket conversation for each case: | Test case | Example | Expected behavior | | ------------------------------- | ----------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------- | | Customer follow-up | The order-tracking agent answers. The customer asks for the tracking link within 30 days. | The same agent continues with the order context when no human replied in between. | | Late follow-up | The customer writes again more than 30 days after the agent finished. | The previous agent does not restart. The message is checked for another eligible automation. | | Human reply after finished work | The agent answers and finishes its current work. A teammate replies next. | That agent does not continue its follow-up chain. The next customer message can be checked for a fresh match. | | Human reply during a pause | The agent asks for a photo and waits. A teammate replies before the customer sends it. | The old agent cannot resume. While the pause remains open, it also blocks a fresh AI Agent. The human team owns the conversation. | | Specialist hand-off | A returns agent learns that the item arrived damaged. | It hands the conversation to one eligible damaged-product specialist, and the returns agent stops. | | Human hand-off | A customer asks for an exception that no specialist can decide. | The agent hands the ticket to the configured person, team, or human queue. | | No action needed | The customer says, "Thanks, that solved it." | The case can end without a reply or human alert when nothing else is needed. | | Messages during work | The customer sends two updates while the agent is still checking an order. | After current work, the agent receives the latest update. No duplicate agent starts. | Also test a later customer message after a human hand-off. It may be matched to an eligible AI Agent again. If the ticket must stay human-only, test after narrowing or disabling matching AI Agent triggers, stopping Classic AI auto-reply, and narrowing or disabling macro Workflows with AI prompts. ## Create useful test cases Use real customer wording when possible. Short synthetic prompts often miss the details that make support cases hard. Define what the agent should do: answer, ask a follow-up, use an action, request approval, or escalate. Add examples where the agent should not act. Check that the agent uses only the actions needed for the case. Rerun the relevant cases after changing instructions, triggers, skills, actions, or approvals. ## Use run history Use it to check: * why this agent received the ticket * which instructions and skills influenced the run * which actions the agent used * whether approval was requested * what customer-facing reply was produced * why the agent escalated or stopped The public [run history link](https://armin.cx/app/_/ai/agents-v2/runs) covers all agents. Agent-specific URLs contain private IDs and should not be published. Open an agent from the [AI Agents overview](https://armin.cx/app/_/ai/agents-v2), then select **Runs** or **Tests**. ## Debug unexpected behavior If a run did not behave as expected, check: * whether another automation already owned the conversation * whether an exact or meaning-based trigger selected a different agent * whether the agent instructions were too broad * whether a skill contained conflicting guidance * whether the needed action was enabled * whether approval blocked the final step * whether the ticket lacked required customer or order context [Ask Torben to diagnose the run](https://armin.cx/app/_/ai/agents-v2?askTorben=Help%20me%20diagnose%20why%20an%20AI%20Agent%20did%20not%20handle%20a%20ticket%20as%20expected.%20Ask%20me%20for%20the%20ticket%20number%2C%20agent%20name%2C%20or%20run%2C%20then%20suggest%20specific%20improvements.). Include the ticket number, agent name, or run so Ask Torben can find the relevant execution. For selection problems, read [Routing and ownership](/ai-agents/routing-and-ownership). ## Best practices * Test before activation and after every major instruction or action change. * Keep approval on while you review early production runs. * Add a negative test whenever the agent overreaches. * Use run history to improve instructions, not just to debug failures.