The Verification Bottleneck

You put AI agents into production. Now you can't watch them.

As agents move from demos to real workflows, the hard problem is no longer whether they are capable enough — it's whether anyone can verify what they're about to do, fast enough, before something irreversible happens. Generation got cheap. Verification didn't. And verification, not capability, is now the real limit on how much autonomy you can safely grant.

The Verification Bottleneck is the field guide to that limit — for the engineers, technical leaders, and governance owners who have to make agents autonomous without owning the blast radius. Drawn from a real oversight system the author built and then attacked, it shows:

• Why "a human reviewed it" quietly becomes a rubber stamp at scale — and the measured cost of handing a person the raw transcript
• Why "just add an AI to check the AI" fails (it detects the danger and mis-rates it) — and what actually works
• An architecture that gates the provably-dangerous on rules an attacker can't talk out of, and escalates only what genuinely needs a human
• How to raise the autonomy ceiling deliberately, one risk class at a time, without trading safety for scale
• How to produce a tamper-evident record of human oversight — the kind the EU AI Act's Article 14 requires from August 2026

It is honest throughout about what it proves and what it doesn't: the results come from a controlled simulation, clearly labeled as such, not a study of human subjects. This is the rare oversight book written by someone who built the thing, measured it, and attacked it — and who tells you exactly where it stops being trustworthy.

If you're responsible for AI agents that act on their own, this is how you keep them on a leash you can see through.

Also from Vita Indarra: Building Reliable AI Agents, Sovereign Machines, and The Glass Box.