Verifier loop

A verifier loop is an agentic design pattern in which verification or grading of work is performed in a separate context window from the context that produced the work, allowing independent, unbiased evaluation within iterative build-check cycles.

Core design

Lance Martin (Anthropic) describes the fundamental structure as maintaining two distinct contexts: a build context that performs the task and a verifier context that evaluates the output — "you can have a loop of a build context and a verifier context." → The key insight is that separation is not merely organizational but functional: having the same context both produce and grade work is less effective than delegating verification to a fresh, independent context window. →

Rationale and effectiveness

Martin argues from empirical observation that "it's quite effective to separate verification into a separate context window," → and that the pattern proves "a very powerful paradigm especially for working with some of the higher capacity models." → The independence of the verifier context prevents the same model from rationalizing or anchoring on its prior outputs when judging correctness — a form of implicit self-grading bias.

Role in long-horizon tasks

The verifier loop is positioned as a component of the broader loop paradigm for long-horizon tasks. Martin frames loops generally — paired with high-capacity frontier models — as "a very good general primitive for long-running asynchronous work," → → with the verifier loop being a concrete instantiation of that primitive suited to tasks requiring quality control over extended, multi-step processes.

Points of disagreement

No speakers in the available material dispute the verifier loop pattern. The sole critical edge in the graph is itself an advocacy position — the claim that same-context verification is inferior — which reinforces rather than challenges the pattern. Broader disagreement about loop-based architectures generally is not recorded here.