Investor's AI Agent Swarm Found $787M in Odd Government Billing

Who is behind the claim

Over the past 10 days, I became the digital Nick Shirley.

Matt Mickiewicz is best known as a co-founder of Hired.com, 99designs.com, and Flippa.com, which gives the announcement a certain weight in startup and investment circles. His profile describes him as an INTJ with a Questioner tendency, and he lists SlashSpend.ai as his own site, framing the experiment as a spending-visibility issue rather than a pure stunt.

The stated goal was a cross-check that is impractical by hand: matching public records against private ones at a scale no small team could manage in the same window. That framing is why the story is being read as a proof of concept for AI-assisted auditing, not just a headline number.

• The operator

  Matt Mickiewicz, co-founder of Hired.com, 99designs.com, and Flippa.com.

• The method

  An AI agent swarm running parallel cross-checks across public and private records.

• The window

  About 10 days from start to the reported finding.

• The claim

  Roughly $787 million in government billing that made no sense.

Why the 10-day window matters more than the number

The eye-catching part is the dollar figure, but the more consequential detail is speed. Ten days is short enough that the same approach could be pointed at other datasets, budgets, or agencies without hiring an army of analysts. For developers and civic-tech builders, that is the real signal: agent swarms are being pitched as a way to compress months of manual reconciliation into days.

It also raises the bar for what counts as evidence. A flagged charge is a lead, not a conviction. False positives, mismatched context, and records that look odd for legitimate reasons are all part of the terrain, so the value of the swarm depends on how well its output can be checked by people.

The verification problem nobody can skip

Because government billing is involved, the announcement alone cannot settle the full picture. The findings have not been independently confirmed, and no public build, download, or official page exists yet, so there is nothing for outsiders to inspect or reproduce. Until that changes, the claim sits in the space between a compelling demo and a verified audit.

That is also where the most useful discussion is happening. Supporters see a template for automating work that humans cannot realistically scale. Skeptics want to know how candidates were ranked, what counted as anomalous, and who signs off before any charge is called wrong. Both reactions point to the same next step: follow-up evidence and third-party review.