Skip to content

The Verifiability Frontier

mihail-eric on where coding agents hit their ceiling in 2026:

"There's definitely stuff and systems that we interact with as software developers when we're creating production code that are not yet verifiable in an easy way. And that's probably where coding agents even today will hit their limits a little bit."

Verifiable tasks (tests pass, schema matches, lint clean) are already agent-territory. Non-verifiable tasks (UX judgement, long-term architectural coherence, production SLO reasoning under novel load patterns) still require a human.

Why this matters

Pairs tightly with verifiable-systems-for-agents — Eric's claim is that the verification layer is now the bottleneck, not the generator. It also connects to Bantilan's durable-observable-debuggable-agents: infrastructure that makes behaviour observable and debuggable is precisely what expands the frontier of what agents can reliably own.

Connects to