About 600 PRs, two weeks, one workday of human time: a long-running loop with a second agent that argues back
Oct 20 Peter Steinberger: a planner-plus-coder loop moved OpenClaw's web frontend from Lit to Solid 2.0 while he checked in every other day. The piece that kept it on track was a separate agent challenging the work, and the misses were the visual ones.
Source: Peter Steinberger on X: moving OpenClaw's web frontend from Lit to Solid 2.0
Peter Steinberger kicked off a loop two weeks ago to move OpenClaw's Lit-based web frontend to Solid 2.0. It took about six days before Lit could be removed and another week to move the code to idiomatic Solid, across roughly 600 pull requests. He "checked in with the session every other day" and calls the process "mostly automatic." Asked how much human effort it took across prompting, review and testing, he answered: "maybe one workday in total."
The setup is small and specific. One model planned, another wrote the code, and the planner launched the Codex command-line tool in separate git worktrees rather than spawning subagents, which he says worked "FAR better." A second agent oversaw the work to keep the planner on track: like a goal file the agent checks itself against, "but instead of the same agent reviewing a text file, a second agent that challenges." Because Solid 2.0 is a release candidate the models barely know, he gave the coder a cheat-sheet skill with the new APIs.
What got through: "a few tiny regressions the agent missed, mostly different shapes on buttons or slight ux regressions." The result, in his words, is a smaller JS and CSS footprint and finer-grained updates.
Why it matters for code review
This is the plainest case yet for steady harnessed throughput over session heroics: a person spending about a day of attention, spread across two weeks of check-ins, while the loop ran on its own. The review work moved into the harness itself. An independent challenger replaced self-grading, worktrees kept parallel changes from colliding, and a context skill closed a knowledge gap the model could not fix by trying harder. It is the same maker/checker split behind loop engineering, run at migration scale.
It also shows where the human still earns their day. Agent review caught drift in the plan; it did not catch a button that changed shape. Visual and UX regressions are the class a diff-reading reviewer is worst at, so a long-running loop still needs a person, or a screenshot check, looking at the product rather than the code. This is one developer's report on one project, not a benchmark.