The standard repair for unreliable agent output is to add a critic. One model produces; another evaluates; the first revises. The pattern is simple, legible, and often useful. It is also easy to fake accidentally.
If the critic receives the producer's framing, reasons with the same model, checks against a vague rubric, and has no authority to reject the artifact, the second call may do little more than restate the first in evaluative language. The system has doubled inference while preserving the original blind spot.
Restless keeps producer and critic as possible organisational responsibilities, not mandatory stages. The critic has to earn its boundary by creating evidence or judgement that the producer could not cheaply supply alone.
Review is not independence
A worker can review its own output and catch obvious defects. Asking it to reread a function after a test fails is cheap and sensible. Self-review becomes weaker when the error is embedded in the worker's original decomposition: it solved the wrong problem cleanly, selected a flattering source, or treated a visual preference as product truth.
A second agent can inherit the same decomposition before it sees the artifact. The producer's explanation tells the critic what to notice and why each choice was reasonable. Shared system prompts and model tendencies add another source of correlation. Two answers are not independent merely because they arrived in separate messages.
Independence is a property of evidence and framing. A browser rendering can contradict code-level confidence. A primary source can contradict a polished summary. A held-back company can reveal vocabulary leakage that the design team no longer notices. A critic is valuable when it introduces one of these genuinely different views.
Start from the outcome, not the rationale
The critic should first receive the intended outcome, accepted constraints, exact candidate, and definition of convincing evidence. This lets it encounter the work as the reviewer or user will. The producer's rationale can remain available for later diagnosis, but it should not pre-empt the critic's first judgement.
This is especially important for design and writing. Explanations are persuasive. A weak visual can sound inevitable once its author describes the conceptual system behind it. Dense copy can sound sophisticated when every sentence is defended individually. Native review asks whether the artifact works before asking why it was made that way.
For code, the equivalent is reproduction. Run the test. Exercise the public interface. Inspect the failure state. A summary of checks is not the check. The critic should be able to falsify the producer's claims using evidence that exists outside the producer's narrative.
A rubric can coordinate blindness
Rubrics make evaluation comparable. They also define what cannot be seen. If every critic scores the same seven dimensions, defects outside those dimensions become invisible by design. A high total can hide one fatal condition, and numerical precision can disguise uncertainty about what the work was meant to achieve.
Restless uses success contracts and evidence grammars to make standards explicit, but they remain contextual. A consequential invariant can fail the artifact regardless of aggregate quality. Open-ended review can name a new material defect and explain why the existing rubric missed it. The evaluation system should learn from repeated omissions rather than treating the rubric as complete.
The strongest critic therefore combines a known contract with permission to challenge the contract. It checks what the organisation expected and asks what an intelligent sceptic would notice that the organisation did not.
The critic needs consequence
Review without the ability to alter the outcome becomes commentary. If a critic can identify unsupported proof but the producer may mark the work complete unchanged, the role exists only for appearance. The review must write to a real decision: accept, reject with exact evidence, request a revision, or escalate a disagreement that depends on owner judgement.
That authority should be narrow. A critic does not silently rewrite the artifact and then approve its own version. It identifies the defect, attaches evidence, and returns responsibility to the producer or accountable lead. This preserves attribution and allows the next attempt to address the actual rejection.
A good rejection changes the work. Restless's early coordination acceptance criterion stated this bluntly: a producer–critic pass had to produce at least one material revision rather than two model calls where the second merely agreed.
One live run shows both value and confounding
In Restless's company-identity dogfood, a control produced a nine-channel asset package for $2.22 and commissioned no independent critic. The first treatment commissioned a critic, which rejected two revisions for invented product proof, overstated review independence, and unsupported voice or authorship. Those were material defects.
But the treatment runtime was stale and had bound company Truth while omitting Voice, Visual Language, and Culture. The comparison therefore could not support a clean causal claim that the critic improved quality. Restless retained the pair as failure evidence rather than reporting an uplift.
After runtime reconciliation and a clean-room reset, the treatment bound all four identity pillars. Three review cycles removed unsupported chronology, a founder placeholder, and first-person-plural interface copy. The exhausted path was replaced by a bounded provenance editor and a fresh critic, which accepted the final exact package. Total treatment spend reached $15.75, including failed branches and repairs.
Different critics find different failures
There is no universal critic. A deterministic test is excellent for an enumerable invariant and useless for tone. A native browser reviewer sees layout and interaction but may miss a false factual premise. A domain expert can judge consequence while overlooking implementation drift. A different model may bring useful diversity or merely a different style preference.
The critic should match the risk. Use reproduction for capability claims, primary sources for factual claims, adversarial tests for security boundaries, native artifacts for experiential quality, and owner review for taste or strategic consequence. Combine critics only when their evidence is meaningfully different.
This avoids the temptation to build a committee around every artifact. Independence has a cost in briefing, reconstruction, latency, and disagreement resolution. The expected defect must be valuable enough to justify that cost.
- Ask what blind spot the critic is meant to introduce.
- Give it direct access to the candidate and relevant source evidence.
- Prevent the producer's rationale from becoming the opening frame by default.
- Require review to change an acceptance decision, an artifact, or an evidence claim.
Agreement can be a warning
Rapid agreement feels efficient. It can mean the artifact is obviously good. It can also mean the roles are insufficiently independent, the rubric is too shallow, or the critic lacks the time and tools to investigate. A system that rewards consensus will learn to produce it.
Disagreement is not automatically valuable either. Critics can manufacture objections, overfit taste, or prolong work beyond economic value. The accountable lead must decide whether the new evidence changes the outcome. The critic owns scrutiny; it does not own the project indefinitely.
Useful review makes the reason for agreement or disagreement inspectable. It points to the artifact, the standard, and the observed consequence. This allows the organisation to distinguish a genuine quality signal from stylistic friction.
The right question is what the second mind can see
Adding a critic is an organisational design decision. It creates another responsibility boundary and spends another context window. That can be justified by independent evidence, a high consequence of false acceptance, specialised expertise, or a producer incentive to stop too early.
It is not justified by a diagram that says every producer needs a reviewer. Sometimes the best critic is a compiler, a customer, a screenshot, or a reconciled provider receipt. Sometimes self-review plus a deterministic gate is enough. Sometimes the owner must see the exact result.
The point is not to make the worker feel watched. It is to make acceptance harder to fool. A critic earns its place when it can see a defect the producing path was structurally unlikely to see—and when the organisation is willing to let that evidence change the work.