AGNTCon + MCPCon Europe 2026

Your coding agent earns trust because it has an oracle: a compiler and tests catch its mistakes. Most agents don't. They pass their evals and still ship confident, wrong answers, because for many tasks, there's nothing to check against.

Two failures hide under "confidently wrong": the model doesn't know (retrieve it), or it knows wrong, a priori it reasserts even against a correct context, which RAG won't fix.

A field guide, three honest moves. Verify, when you can check: deterministic checks on the answer, not the tool calls, run in the harness. Abstain, when you can't: make the agent say "I can't verify this," gated on an external check, not its own confidence. Amplify, when there's no right answer, only judgment: the model collapses toward the average exactly when you want the opposite, so you push it off and amplify the novelty you put in, judged by a human.

We ground all three in our own builds where the model is reliably wrong: a finance agent, a genomics agent that invents false-but-plausible mechanisms, and a microtonal-music agent that keeps dragging toward Western tonality. The move comes down to two questions: can you check the answer, and is there a right one?