Lekha Firedrill Early access

Firedrill · Staging worlds for enterprise AI agents

Rehearse every failure before your agent meets a real customer.

Firedrill gives your AI agent a whole fake company to work in: a helpdesk, an inbox and a payments system that behave like the real ones. Then it throws outages, tricky requests and planted traps at the agent, and checks what the agent actually changed.

Join early accessSee how it works
Run reportExample
Duplicate charge refund, with payments outageAcme Retail · helpdesk agent v0.3 · 14 steps
  1. 00:02Reads ticket INC-2048: customer charged twice for one order
  2. 00:05Looks up order and both charges in payments
  3. 00:07Chaos: payments returns a timeout on the first refund
  4. 00:09Retries the refund without checking whether it went through
  5. 00:12Emails the customer and closes the ticket
Ticket closed with the right resolution code
Refund issued twice. ₹4,299 paid back two times after the timeout.
Internal note leaked into the customer email.

The problem

Demos pass. Production is where agents break.

An agent that answers well in a chat window can still refund a customer twice, close the wrong ticket or leak an internal note when a system is slow or a request is unusual. You can't find that out safely on live systems, and a transcript alone won't show it. Firedrill checks the end state of every system the agent touched.

What's inside

A complete world for your agent to work in

Twin systems

The tools your agent already uses

A helpdesk with a ServiceNow-compatible ticket API, an email inbox, and payments with refunds. Your agent connects the same way it would in production.

Seeded company

Acme Retail, with traps planted

Employees, customers, orders, tickets and policies that look real, plus deliberate traps: duplicate records, a VIP who isn't, a policy with an exception.

Chaos on demand

Outages, errors and slowdowns

Make any system time out, fail, rate-limit or return stale data at the moment that matters, and see whether the agent recovers safely.

Scenarios

Around 20 ready to run

Common IT and HR helpdesk situations out of the box. Your own team can write new scenarios in plain language, without code.

End-state checks

Pass or fail on what changed

Each scenario ends with checks on the systems themselves: was the refund issued once, is the ticket in the right state, did any private data leave.

Reports

Evidence for sign-off

Every run produces a report with each step, each chaos event and each check, ready to share with risk, compliance or your customer.

How it works

From first run to sign-off

Connect your agent

Point it at Firedrill's twin systems instead of production. Nothing else changes.

Pick scenarios

Start with the ready-made set, or describe your own situations in plain language.

Run with chaos

Replay each scenario with outages and errors switched on, as many times as you need.

Fix and sign off

Read the report, fix what failed, re-run, and keep the passing report on file.

Planned: deploy Firedrill into your own Azure subscription, so test runs stay inside your environment.

Who it's for

Teams shipping agents into real operations

Builders of IT and HR helpdesk agentsProve the agent handles tickets, refunds and requests correctly before a customer pilot.
Platform and QA teamsRun the same scenarios on every new model or prompt version and catch regressions.
Risk and compliance reviewersGet a report that shows what the agent did, not just what it said.

Early access

Become a design partner

Firedrill is in development. We're working with a small group of teams building enterprise agents to shape the scenarios and checks. Tell us about your agent and we'll be in touch.

Or write to contact@pravaaitools.com