Skip to content

Humanize 1: RLCR Loop

The one everything else grew out of. One agent builds, a second arrives with no memory of how the work was done and reviews it, and the findings become the next prompt — until the acceptance criteria are met, not until the model says it is finished.

PolyArch/humanize · proposed at UCLA PolyArch · derived from GAAC · MIT

What RLCR stands for

Ralph Loop with Codex Review — the ralph-loop plugin with an independent Codex review in every round. It also reads as Reinforcement Learning with Code Review, which is what the loop is actually doing: the review is the reward signal, and the next round is the update.

What it is made of

Iteration over perfection

Nobody gets it right in one shot

The work is refined in rounds instead of one long prompt hoping for a finished answer. A problem found in round two costs a round, not a rewrite.

One build, one review

The builder is not the judge

Claude implements; Codex reviews and marks severity. Two models with different blind spots, and neither grading its own homework.

The loop

It ends on the criteria, not the vibe

Findings feed back into implementation until every acceptance criterion is met. Optionally in parallel, with agent teams working the same plan.

Begin with the end in mind

The person is checked too

Before the loop starts, it checks that you understand the plan about to run. You stay the architect; the agents are the leverage.

Installing it

It is a Claude Code plugin, so it installs from inside Claude Code:

/plugin marketplace add PolyArch/humanize
/plugin install humanize@PolyArch

Reviews go through the codex CLI, which has to be on your path. Add #dev to the marketplace line for the development branch.

The commands

CommandWhat it does
/humanize:gen-ideaOpens a loose thought into a draft, exploring several directions at once
/humanize:gen-planTurns a draft into a plan, written to a file you can read and edit
/humanize:refine-planFolds a reviewer's inline comments back into the plan before any code is written
/humanize:start-rlcr-loopRuns the loop: Claude implements, Codex reviews, repeat until the criteria are met
/humanize:ask-geminiHands a research question to the Gemini CLI when the answer is on the web

A humanize monitor dashboard runs in a second terminal, so a loop left going for hours is something you can watch rather than something you find out about afterwards.

What became of it

The method moved into the flowverse as the humanize1 flows — the same commands, each on its own agents — where it sits beside RLAR, Flame Chase and the Ralph loop and can be run, forked and beaten like any other flow. It set the rule the rest are built on: the reviewer arrives with no memory, reads the repository rather than the transcript, and what it noticed is what the actor hears next.

The loop outgrew being a plugin. A plugin can drive the coding agent it is installed in; Humanize 2 drives ten of them, on machines that are not yours, for eleven hours at a time, and writes down everything that happened.

Humanize 2: Agent Flow System · FlowBench · who built it