A Field Guide to Confidently-Wrong AI Agents
Michał Orzechowski heads to AGNTCon + MCPCon Europe with a practical playbook for building AI agents you can actually trust.
Sano researcher Michał Orzechowski will take the stage at AGNTCon + MCPCon Europe in Amsterdam on Thursday, 17 September 2026, with a talk titled “Verify, Abstain, or Amplify: A Field Guide to Confidently-Wrong Agents.”
A coding agent earns our trust because a compiler and a test suite catch its mistakes. Most AI agents have no such oracle. They pass their evaluations and still ship confident, wrong answers — because for many real tasks, there is simply nothing to check against.
Orzechowski’s talk offers three honest moves.
- Verify: when you can check the answer, run deterministic checks on the output itself, inside the harness.
- Abstain: when you can’t, make the agent say “I can’t verify this” — gated on an external check rather than the model’s own confidence.
- Amplify: when there is no single right answer, only judgment, push the model off the bland average it drifts toward and amplify the novelty you feed it, with a human as the judge.
Each move is grounded in real systems where models are reliably wrong: a finance agent, a genomics agent that invents false-but-plausible mechanisms, and a microtonal-music agent that keeps sliding back toward Western tonality. In the end, it comes down to two questions: can you check the answer, and is there a right one?
The talk
“Verify, Abstain, or Amplify: A Field Guide to Confidently-Wrong Agents” Michał Orzechowski
Thursday, 17 September 2026 · AGNTCon + MCPCon Europe, RAI Amsterdam (17–18 September 2026)
Schedule: bit.ly
