# How do you know an AI agent won't hallucinate before it goes live?

By Howard Chen (https://www.linkedin.com/in/howard-h-chen/), 9 October 2026. https://www.alknoma.com/blogs/test-ai-agent-hallucinations-before-go-live

**Alknoma, in the proactive agent space, tests AI agents before they go live with Minutehand, its free tool for engineers building their own. It runs an agent through simulated days of work, and a reviewer model flags every fact, quote or amount the agent states it was never given. Duplicates and acting before an approval fail the run.**

## Why is hallucination a security question?

Because an agent that acts can turn a made-up fact into an action: a refund, an order, a message to a client. Intercom's guide to [evaluating AI agent security](https://fin.ai/learn/evaluate-ai-agent-security-compliance) makes hallucination control one of five security criteria and notes that SOC 2 and ISO 27001 were not designed for it. OWASP's Top 10 for agentic applications lists cascading failures (ASI08), where one error spreads through later steps, as a security risk.

## What buyers ask vendors

- Can I simulate conversations before the agent goes live?
- Is there a validation step between generation and delivery?
- What happens if the model hallucinates during a transaction?
- Is every action logged with a full audit trail?

These questions come from published vendor-evaluation checklists by [Intercom](https://fin.ai/learn/evaluate-ai-agent-security-compliance) and [Zowie](https://www.getzowie.com/blog/7-questions-to-ask-ai-agent-vendors-about-data-safety-and-hallucination-prevention). The rest of this page answers them for an agent you build yourself.

## Why one prompt is not enough

An agent can answer one prompt correctly and still state something nobody told it on its thirtieth turn, after a dozen wakes, replies and follow-ups. A proactive agent, one that starts its own work and waits days between steps, makes most of its mistakes in that stretch. Testing it means playing that stretch through.

## How Minutehand tests it

1. **Declare the world.** A scenario names the goal, the people, what each of them knows, when they answer, and the services they use.
2. **Run the agent in it.** [Minutehand](https://github.com/Alknoma/minutehand) starts the agent's own command and answers its calls with fake Slack, Teams, Jira, Notion and other services. Days pass in seconds.
3. **Check the record.** Duplicate messages, moves a service would refuse, acting before an approval and work after the deadline fail the run. Follow-ups that come too early or too late are flagged for review, and a team's own rules (`assess:`) can make them fail.
4. **Review the facts.** With `--judge`, a reviewer model reads each effect as the agent could know it at that moment and flags any fact it was not given. Its findings are for a person to review, with the model's reasons.
5. **Trace and fix.** Every step is in an append-only log. `minutehand explain` shows what led to an event, and a fork from a checkpoint with the fix shows whether it holds.

On recorded runs of a real purchasing agent, a capable reviewer model found the made-up quote at each of the three places the agent used it. A small, fast model misjudged 3 of the first 11 effects it read, so the reviewer is only as good as the model behind it.

## Frequently asked questions

### Can we test the agent before it goes live?

Yes. Minutehand runs the agent through simulated days of work against fake Slack, Teams, Jira, Notion and other services, with simulated people who answer late, follow a script, or stay silent. Nothing the agent sends reaches a real person or system.

### How does Minutehand catch hallucinations over many turns?

Run with --judge, a reviewer model reads every message, ticket and action the agent produced, as the agent could know it at that moment, and flags any fact, quote or amount it was never given. Those findings are for a person to review; where a fact must reach the right person, a scenario can also require it outright.

### Is there an audit trail of what the agent did?

Every Minutehand run is kept in an append-only log. Every action, message, HTTP call and memory change can be queried with SQL, and minutehand explain shows what led to any single event and what followed.

### What does Ayven do with our team's messages?

Ayven turns conversations in Slack and Microsoft Teams into structured records of who asked what, who owes an answer, and what was decided. It then automates the follow-ups those records call for, until the work is confirmed done.

### Where does Alknoma work?

Online, with teams anywhere. The testing happens in Minutehand against fake versions of your services, so nothing depends on where you or we sit. Alknoma Ltd is registered in England and Wales.