Module 3 · Planning, Context & Multi-agent Systems · scripted
The Actor–Critic
The loop that grades itself
The last archetype attacks a different weakness. Everything we've built produces artifacts — reports, plans, code, designs — and a first draft of anything is rarely right. The fix is a loop older than software: one party produces, an independent party criticizes, the producer repairs — until the work passes. The name comes from a long tradition (reinforcement learning calls its version actor–critic), and the theater thread makes it feel inevitable: every play needs its critic.
The actor–critic loop
One classification before we run it, because it marks a border in the archetype family. The Explorer and the Scientist live in a single conversation — one wall, one trajectory. The Actor–Critic is our first multi-conversation archetype: two minds that never share a wall, coupled through a third thing — the artifact. The actor builds and modifies it; the critic observes and critiques it; it is the only object they have in common. And the reason for two minds is not redundancy — it is that building and checking are different attentions:
Two attentions, one artifact
THE ACTOR builds. Its attention lives inside the work — making choices, holding structure, moving forward. Building attention is forward-looking, and it reliably misses things: a builder reads what they MEANT, not what is on the page. THE CRITIC observes. Its attention has exactly one job: completeness and correctness against the bar. It builds nothing, wants nothing, moves nothing forward. Between them sits the ARTIFACT — written by one attention, read by the other. Neither attention can do both jobs at once; that is why there are two seats.
Now watch it run.
The grant package
The task: assemble a grant proposal package — narrative, budget, references — under a funder's hard rules and soft expectations.
The critic's bar — layered like the receipt validator
PASS = all code checks green AND both judge checks YES.
[the critic reviews actor draft v1]
FINDINGS (3):
- BUDGET — columns sum to 2,847,240 kr but the total row says 2,847,000. WHY: the equipment subtotal was rounded. NEXT: set total = 2,847,240, or re-round line items per funder rule §4.2.
- CITATION — key [Moen2024] does not resolve in references. WHY: reference list has "Moen2023". NEXT: correct the year in whichever is wrong; verify against the actual paper.
- CRITERION 2 — the evaluation section measures only technical milestones. WHY: funder's criterion 2 requires societal-impact evaluation; §3 never mentions the stakeholder outcomes promised in §1. NEXT: add an evaluation instrument for the §1 stakeholder claims — survey, adoption metric, or case study.
[the critic reviews actor draft v2]
Code checks: all green. Criterion-2 check: §3 now includes a stakeholder-adoption study tied to the §1 claims — YES. Consistency check: YES. PASS. One advisory, not blocking: the new study adds 40,000 kr not yet in the budget narrative — flagged for the PI.
Why the critic must be independent
Here is the deepest design rule of the archetype, and it comes straight from the self-conditioning thread. Suppose the actor checks its own draft, in its own conversation. That wall contains every reason the draft looks the way it does — each decision, already justified, each phrasing already approved once by the very trajectory that produced it. Asking that conversation "is this good?" invites the answer the trajectory has been building toward all along: yes.
The independence rule
The actor cannot grade its own homework on its own wall. A critic must be a FRESH conversation: the draft, the bar, and nothing else. It has never met the reasons. It reads what is actually on the page — which is exactly what the funder will do. (Fresh eyes are a firebreak against self-agreement — the same reason the compaction scribe and the ask-the- document tool were disposable conversations.)
There's a second, quieter benefit: the critic's wall is small — the draft and the bar, no history — so its judgment runs at full attention. Producer walls bloat; verifier walls are born clean, every round.
The bar and the budget
Two contracts keep the loop honest, both written before it starts:
The two contracts
Ship · revise · escalate: the critic's verdict is a label, and a label can be routed by code. The actor–critic loop is itself just a labeled workflow — minds in both steps, code counting the rounds.
That last observation ties the room together: the archetypes compose. An actor–critic loop can sit as one stage inside a Planner–Executor's plan; the Designer's simulate step is a code critic; the Scientist's discriminating test is criticism aimed at hypotheses. The four archetypes aren't rival architectures — they are shapes that nest.
Where the actor–critic earns its cost
The loop doubles LLM spend per round — producer and critic both think. It pays for itself where:
Choose this archetype when
Where there is no articulable bar, a critic degenerates into taste — and taste belongs in the actor's brief (preferred and non-preferred examples), not in a loop that can't converge.
- Take finding 3 from Example 1. Write the judge check that caught it as an actual prompt — what does it receive, what question is it asked, what must it return?
- Your field has a real critic: the reviewer. Which parts of their review are code checks in disguise? Which are judge checks? Which are taste — and what does your answer imply about automating any of it?
Take an artifact type you produce repeatedly — papers, proposals, analysis reports, code. (1) Write the bar: every code check as a rule, every judge check as a verbatim one-thought prompt. (2) Set the budget and the escalation. (3) Run one real artifact of yours through the loop on paper: play the critic honestly, write the findings in WHAT/WHY/NEXT. (4) The finding that survived two rounds — is it a repair problem, a bar problem, or an actor-brief problem? Fix it at the right layer.