Why every Claw ships with exactly one Critic, no matter what it does

Claws enforces a different shape. Every Claw, official or custom, is built from the same three roles. That's not a suggestion: the visual builder won't let you deploy without all three.

The three roles

• The Orchestrator (exactly one) is the only agent that talks to you. It reads what's due, delegates tasks to the right Workers, collects their output, and sends it for review before anything reaches you. It runs on a fast, cheap model, since routing doesn't need deep judgment.

• Workers (2 to 6) are the specialists: research, writing, analysis, outreach, whatever the Claw is built to do. Most run on a mid-tier model; creative roles often run on the strongest one instead. Each has its own instructions and its own job. You never see their false starts, only what survives review.

• The Critic (exactly one) reviews everything before it ships, and runs on the strongest available model, since judgment matters more here than anywhere else in the team. Every Worker output gets scored against a quality threshold you can set, 8 out of 10 by default. Score below the bar, and the work goes back with specific feedback. Score above it, and the Orchestrator presents it to you.

What the Critic actually checks

The default review runs against four criteria:

• Accuracy: facts, numbers, and names correct.

• Completeness: fully addresses the task, no gaps.

• Quality: professional, well-structured.

• Risk: nothing that could create a legal, compliance, or reputational problem.

Each Claw's Critic can add domain-specific checks on top: a Fair Housing check for real estate communications, a HIPAA-aware gate for medical intake, a bias check for recruiting outreach. A 6 out of 10 is a reject. Only 8 and above passes.

What happens when work fails review

The Critic's feedback goes back to the Worker, who revises. This can repeat up to three rounds. If the work still isn't clearing the bar after three attempts, the Orchestrator presents the best version to you directly, flagged as below threshold, rather than looping forever or quietly lowering the standard.

You always see the score, and you always see why something was rejected, if it was. Review happens before anything reaches you, every time. At launch, how much an agent can act on without asking again will be graduated based on what you've approved before (more on that in a separate post).

Why this is architecture, not a feature

You could build a single do-everything agent. It would be simpler, and cheaper to run, too. It would also have no internal check on its own output. The same model that wrote the draft would have to judge it too, and that's a conflict most people wouldn't accept from a human employee, let alone from software making decisions for their business.

Splitting the roles means the Critic never wrote what it's reviewing. It has no incentive to wave through weak work. That separation is the entire point.

Every review gets logged too, not just the final score. If you need to see why something passed or failed six months from now, the trail is still there, timestamped, tied to the agent that made the call.

Every Claw, from the free templates to the largest official one, runs this same pattern, permanently, with no setting that turns it off. That permanence is what makes autonomy something you can actually trust with your name on it.