Skip to content

// Feedback loops · ~3 min read

Critic-Actor

Reach for it when: a dedicated critic reviews every actor draft before it ships.

After this lesson you can: Pair a drafter with a separate critic that returns pass or revise, with a retry cap and a hand-off to a human.

// 60-second mental model

How to hold it in your head

An actor drafts. A separate critic scores the draft against a rubric and may request a revise. The critic does not write the piece — it flags. A human (or a hard stop) still owns the ship decision.

Critic-Actor pairs a generator with a dedicated runtime scorer that returns pass/revise (and structured notes). Bounded actor retries may follow. The critic checks one draft at a time, while the system runs.

How this differs

Also known as evaluator-optimizer (Anthropic’s name for this kind of loop) or generate-and-review, which Google ADK also calls generator-critic. This course separates the two by emphasis: a pass/revise verdict and a human hand-off here, a numeric threshold in Evaluator-Optimizer. It is not the actor-critic method from reinforcement learning, where the critic estimates value to train the actor. Reflection: the same model critiques its own draft against a rubric. Evaluator-Optimizer: a separate scorer returns a number, and you revise until it clears a threshold. Critic-Actor: a separate critic returns pass or revise with notes, and after a capped number of retries a human decides.

// Mini architecture

Actor drafts · Critic scores · bounded revise

An Actor writes draft N. A Critic, using a rubric, returns pass, or revise plus notes. If it says revise and retries are left, the Actor revises using the notes as data; if not, the draft is held for a human or the loop stops.

Critic notes are data for the actor, not new system instructions. Pass ≠ auto-publish unless you explicitly wire that.

// Mini-project

Essay draft with rubric critic (stub)

Actor + critic stubs

Goal: Produce a short essay stub; critic returns pass|revise against a fixed rubric; allow at most two actor revises.

  1. Actor stub emits a canned draft with one known weakness (weak hook).
  2. Critic stub scores against a tiny rubric (hook, one idea, honesty) → pass|revise + notes[].
  3. On revise, give the notes to the actor as quoted data to read — not as new instructions.
  4. Cap retries at 2; if still revise, stop with status `needs_human`.
  5. Log actor → critic → retry_count each round.
  6. Assert the critic never rewrites the draft body itself.

A real example from this site’s code: a Writer agent drafts essays and a Critic agent scores each draft on a 7-axis rubric, returning pass or revise. The code is built so that, on revise, the Critic sends its notes back to the Writer automatically, at most twice; after that the draft waits for Mir. Whatever happens, only Mir’s Approve click publishes it.

// Common failure

What goes wrong

Symptom

Critic writes the draft — the actor becomes decorative, or “pass” silently ships without a human gate you meant to keep.

Fix

Restrict the critic to score + notes tools only. Treat pass as “ready for review,” not “published,” unless a separate HITL gate says otherwise.

// Self-reflection

Sit with this

Where does your generator still grade its own homework?

Sources

Session only. Nothing is saved.