critic agent

The boundary is the binding power. A critic agent and a gatekeeper agent get swapped for each other constantly, and both can sit at the same handoff, but only the gatekeeper's verdict can stop the run. A critic raises objections; another step decides whether those objections were resolved. A gatekeeper decides continuation outright and cannot be argued past.

The other thing that makes a critic a critic is coldness. It works without the reasoning trace that produced the artifact, so its objections address what exists rather than what someone meant.

What it catches, and when

Shift-left quality in an agentic dev pipeline shows the cost of late review. Cold inspection after a long chain forces rework of every downstream step built on the flaw. A concrete catch from that pipeline: a public-facing assessment-intake endpoint able to trigger spend, flagged with no rate-limit, budget-cap, or length-guard decisions pinned. That one landed in pre-build review, before any code existed, which is the cheap version. The essay notes that the Critic did not fail; the timing did.

The role travels under two names. Who checks the checker calls it the Challenger. Memory is easy, retrieval is hard describes it as a fresh-context critic that surfaced structural issues yet missed a behavioral one.

Why a critic is not enough by itself

On a 109 KB spec and a 94 KB prompt set, the Challenger passed the prompts, and a focused verification pass reported zero blockers. A delegate then kept going and caught twenty build-breaking defects across three rounds. The Challenger had passed every one of them.

The conclusion drawn there is not that critics are useless. A second reviewer earns its place when it checks how the first reviewer's findings were handled, not when it re-examines the same artifact. That second job belongs to a gate.