Steinberger says Astra running codex cli in worktrees beat subagents

Peter Steinberger's agent loop has finished moving OpenClaw's Lit-based web frontend to Solid 2.0, and he has posted the shape of the run. It took roughly six days until Lit could be removed and another week to move the patterns to idiomatic Solid, about 600 PRs in total. He checked in with the session every other day, and says the process was mostly automatic. The migration itself was earlier reported as it ran.

Peter Steinberger 🦞

@steipete

X

Technical dept is can be counted in token now.

Peter Steinberger 🦞

@steipete

X

prompting Astra to run codex cli in worktrees works FAR better than relying on subagents

The result he reports is a smaller JS and CSS footprint and much more fine-grained update control, with everything feeling more snappy and responsive. The agent missed a few tiny regressions, mostly different shapes on buttons or slight UX regressions. His summary of the whole thing is that "Technical dept is can be counted in token now."

The setup is the copyable part

Two models split the job. Astra did the planning, Sol 6.1 did the coding.

The detail worth stealing is how the planner was driven. Steinberger says prompting Astra to run codex cli in worktrees worked far better than relying on subagents. Worktrees are git's way of having several checkouts of one repo side by side, so independent agent runs do not fight over the same working directory. Subagents, by contrast, live inside the parent session and share its context.

He also added a reviewer. A second agent oversaw the work to keep Astra on track, which he compares to a /goal command, except that instead of the same agent re-reading its own text file, a different agent challenges it. He wrote an OpenClaw plugin for this, calling it a "smart goal" feature.

This is one engineer on one codebase, not a benchmark, and the codebase is his own project. What it does show is a two week unattended run where the human intervention was a check-in every other day, and where the failure mode that survived was cosmetic rather than structural.