Explainer Randy's AI Frameworks

Proposer/Critic — How You Apply the Pattern

One component proposes, a second evaluates before it stands. Where the pattern runs across your work, from ABOM's organization-scale judgment layer to the checker agents in AI Agent for PM.

Generated
November 18, 2026, 3:38 PM MT
Built from
  • wiki/concepts/proposer-critic.md
  • wiki/sources/ai-pm-agent-architecture-v1-2.md
  • wiki/sources/langgraph-agentic-qa-pipeline-notes.md
  • wiki/sources/welcome-to-gas-city-notes.md
  • wiki/sources/pm-decision-framework.md
  • wiki/sources/io-world-github.md
  • wiki/sources/journeycapture-project-notes.md
  • wiki/concepts/abom.md

An agent pattern where one component proposes an action or judgment and a second evaluates it before it stands. The pattern itself is general. What you treat as original is using it at organization scale in ABOM: verifying that agent nodes judge well over time, rather than inspecting every handoff.

ProposerDoes the work: an action, a merge, a problem definition, a judgment.
→
CriticReads the source evidence, not the proposer's reasoning.
ConfirmFlagReject
→
Stands, or goes to reviewNever rewritten by the critic.
The shape as written into AI Agent for PM v1.2's two-agent rule.
Keep judgment and observability apart

"Judgment layer" (Proposer/Critic: is it right?) is distinct from "observability layer" (health signals: is it alive?). ABOM's notes warn that conflating them loses the differentiator.

2scales: organization and system
6places it runs in your work
3critic verdicts: confirm, flag, reject
3orchestration patterns compared

Two scales

Organization scale · ABOM

The mechanism for verifying that business-unit agent nodes judge well over time (calibration, evals, golden cases) rather than inspecting every handoff. You treat it as ABOM's original contribution, with no clean analog in competing frameworks or in OOP.

System scale · LangGraph agentic QA pipeline

The dual-loop architecture of the LangGraph GUI-testing agent, alongside a rule-based Observer and a Reviewer that refines task steps. One of three orchestration patterns you compared, with ReAct and Plan-and-Execute.

Where it runs

WhereProposerCriticStatus
ABOMBusiness-unit agent nodesCalibration, evals, golden cases on how well each node judgesFramework (v2); no running deployment
LangGraph agentic QA pipelinePlanner (plan-guided ReAct)Dual loop with a Reviewer; rule-based Observer flags symptomsBuilt at Keysight; which agent plays which role isn't spelled out in the notes
JourneyDriveAgent acting from screenshotsSeparate reviewer confirms each step from the screenShips an act-then-review loop (repo, Sep 22)
AI Agent for PM v1.2Merge step; Problem Definition agentMerge Checker; Problem Definition CheckerStanding design rule
PM decision frameworkThe call and its fixRisk + Fix: would the fix have caught the failure already in the data?Applied to product calls
Gas City (Yegge)Agent that recommendsSecond agent that checks and actsExternal; your article names a Problem Definition checker as the next place to apply it

The two-agent rule in AI Agent for PM

Any LLM step whose error propagates unseen gets a checker.

One of v1.2's five design principles, added in your 2026-10-20 critical review.

Merge Checker Agent

Re-reads the records behind each same_issue_as edge; confirms, downgrades or rejects.

Problem Definition Checker Agent

Verifies claims against cited records, checks one problem vs. a split, and checks the persona match.

Evidence, not reasoning

Checkers read source evidence, never the worker's reasoning, and can confirm, flag or reject but never rewrite.

Review gates

Checker flags are deterministic triggers for early human review; the worker's own confidence is one input, never the only one.

Open questions