AI agent testing before your customers do it for you

Your agent passed the demo. Demos are polite.
Real users paste weird documents, ask the wrong question and try to make it do things it shouldn't.

I run about 200 of those cases against your agent and tell you exactly where it breaks. 5 working days. £750 fixed.

Book a 20 min call See the report outline

Why AI agents fail after launch

The model did not get worse. The inputs got real.

A customer uploads a PDF with instructions hidden in it. The agent follows them.
Someone upgrades the model or edits the prompt. Three things that worked now don't. Nobody notices for a week.
The agent calls the right tool with the wrong arguments, confidently.
It refuses a fair request and answers one it should have declined.

Each one is a support ticket, a lost deal or a screenshot on X. Finding them before a launch, a big customer or a model upgrade is cheaper.

What the AI agent audit tests

Prompt injection testing

Direct and indirect injection: instructions hidden in documents, tool outputs and user content.

Hallucination and grounding

Answers checked against your sources. Made-up facts, wrong citations, confident guesses.

Tool misuse

Wrong tool, wrong arguments, unsafe actions, loops, actions taken without confirmation. What it did, not only what it said.

Edge and malformed inputs

Empty, huge, multilingual, badly formatted and contradictory inputs.

Refusal and safety

Over-refusal of fair requests, under-refusal of ones it should decline.

Regression across versions

Same suite on two model or prompt versions, so you see what got worse.

What you get in 5 days

A scored report: pass rate per category, severity per failure, a go or fix-first call.
Every failing case, reproducible: input, context, output, why it failed.
Fixes ranked by impact and effort.
The test suite (CSV/JSON). Rerun it on every release.
A 30 min walkthrough.

You stop arguing about whether the agent is "good enough". You have a number and a list.

How it works

Day 0: 20 min call

Scope one agent or feature. Agree access (staging endpoint, API key or test account) and what "correct" means.

Days 1 to 4: testing

I build and run the suite. No production data needed.

Day 5: report

Report, failing cases, fixes and the reusable suite, plus a walkthrough.

Price

£750

Fixed price (+VAT if applicable). One agent or LLM feature, about 200 test cases, 5 working days.

£375

Pilot, first 2 customers only. Same audit at half price, in exchange for a short testimonial if you found it useful.

Ask about a pilot slot

Sample report outline

Summary: overall score, top 5 risks, go or fix-first recommendation
Scope and setup: agent, model and prompt version, tools, test environment
Scores by category: injection, grounding, tool use, edge inputs, refusal and safety, regression
Failing cases: input, expected, actual, severity, how to reproduce
Recommended fixes: ranked by impact and effort
Regression comparison (if two versions supplied)
Appendix: full test suite and method

Who it is for

Seed to Series B teams shipping a customer-facing agent, copilot, support bot or AI SDR, who want an independent check before a launch, a big customer or a model upgrade.

AI agent testing: questions

What is AI agent testing?

Running your agent against a planned set of inputs and checking what it says and what it does: the answers, the tool calls, the actions. Then scoring the failures so you know which to fix first.

How is this different from AI red teaming?

Red teaming hunts for attacks. This audit covers attacks too (prompt injection, tool misuse), but also the boring failures that lose customers: wrong answers, ignored instructions, regressions after a model upgrade.

Do you need production access?

No. A staging endpoint, API key or test account is enough. No production data needed.

How long does the audit take?

5 working days from kick-off to report and walkthrough.

Can we rerun the tests ourselves?

Yes. You keep the full suite as CSV/JSON and can run it on every release.

Will you sign an NDA?

Yes. Your prompts, data and results are not shared.

What does it cost?

£750 fixed price (+VAT if applicable) for one agent or LLM feature. The first 2 customers get a pilot at £375 in exchange for a short testimonial if it was useful.

Book a call

Book a 20 min call or a pilot slot

Tell me what the agent does and when it ships. I reply within one working day.