Someone raised $28M to break AI agents on purpose

On 25 June 2026, a company called Coval announced a $28 million Series A. Norwest led it. Base10 Partners, Twilio Ventures and Y Combinator joined.

The product does one thing. It breaks AI voice agents on purpose, before real customers get the chance.

Coval throws simulated calls at an agent. Accents it has not heard. Interruptions. Background noise. Requests nobody wrote a script for. Failed calls get fed back into the next round of tests automatically.

I build voice and CRM automation for a living. My first thought was not about voice agents.

It was about how few founders ever test their raise.

What got funded was not intelligence

The agents Coval tests already work. That is the part people miss.

Models stopped being the bottleneck a while ago. Any competent team can stand up a voice agent that handles a clean call.

The money went somewhere less glamorous. It went to the catalogue of ways the thing falls over.

That is a real product because failure is expensive and invisible. An agent fails quietly, on one call, to one customer, at 3pm on a Tuesday. Nobody files a ticket. The revenue just does not arrive.

Your raise is a system with no test suite

A raise has the same shape. It is a repeatable process with a lot of surface area and almost no instrumentation.

You rehearse the pitch. Who do you rehearse it with?

Your co-founder. A friendly angel. An advisor who already believes the story. People who fill in the gaps for you, because they know what you meant.

None of them are simulating the call that goes wrong.

The mechanism worth stealing

Coval does three things in order. None of them need a $28 million round.

The version you can run this week

Here is the prompt I use before a founder goes into a partner meeting. It runs step one on your one-liner.

It is deliberately unkind. That is the point.

You are a panel of 12 investors reviewing one company. Each of you has a
different mandate: pre-seed generalist, seed SaaS specialist, deep-tech
fund, fintech fund, growth fund, corporate VC, family office, angel
syndicate lead, emerging-market fund, impact fund, a partner who passed
on this category twice, and an associate screening 40 decks this week.

Here is the company:
[PASTE YOUR ONE-LINER AND THREE METRICS]

Do this:
1. Each investor writes their single strongest objection. One sentence.
   Write it the way they would say it to a colleague, not to me.
2. Mark each objection as one of: story problem, evidence problem,
   market problem, team problem, timing problem.
3. Rank all 12 objections by how many of the panel would independently
   raise the same one.
4. For the top 3 only, tell me what evidence would remove the objection.
   Be specific about the artefact: a number, a document, a customer name.
5. Tell me which objection I cannot fix before this raise. Say so plainly.

Do not suggest wording changes. Do not be encouraging.

The output that matters is step five. Every raise has one objection you cannot answer this quarter. Knowing which one it is changes who you talk to, and in what order.

What to do with the answers

Run it once. Save the output. Run it again after the first five real meetings.

Compare the two lists. The objections the model predicted and real investors also raised are structural. Fix those.

The objections only the model raised are noise. Ignore them.

The ones only real investors raised are the interesting ones. Something about how you tell the story is creating a problem that does not exist in the business.

That gap is the whole exercise. Coval calls it a failure catalogue. On a raise, it is just the list of things you keep getting wrong in rooms you do not get back.

More prompts and templates collect on the resources page. All free, no signup.