Steinberger says worktrees beat subagents on his rewrite

Peter Steinberger has posted the machinery behind his OpenClaw frontend rewrite, and the headline is not the framework. It is the orchestration. Astra did the planning, Sol 6.1 did the coding, and rather than handing work to subagents he had Astra run codex cli in git worktrees instead.

That migration moved OpenClaw's Lit-based web frontend to Solid 2.0 in about 600 PRs, roughly six days to remove Lit and another week to move patterns to idiomatic Solid. We covered the migration itself earlier. What is new is the shape of the loop he ran it in.

Peter Steinberger 🦞

@steipete

X

I found that the workflow of prompting Astra to run codex cli in worktrees works FAR better than relying on subagents.

Peter Steinberger 🦞

@steipete

X

A second agent oversaw the work to keep Astra on track.

A second agent whose job is to argue

Steinberger says a second agent oversaw the work to keep Astra on track, and he is explicit that this is not the same as an agent reviewing its own notes. He compares it to the /goal pattern, except instead of one agent re-reading a text file it is a separate agent that challenges. He wrote an OpenClaw plugin for this, calling it a smart goal feature.

He checked in with the session every other day and says the process was mostly automatic. The regressions that slipped through were small, mostly different shapes on buttons or slight UX problems. His read on the result is that smaller JS and CSS and finer-grained update control made everything feel snappier.

The same loop, pointed at the TypeScript compiler

Separately, Steinberger says he moved OpenClaw to tsrs over a weekend, describing it as a drop-in replacement that is around 5x faster for OC. He notes it would have been 9x faster but they were already on the 7.1-dev version of the TypeScript compiler, and points out that Microsoft took a team and two years to rewrite the TS compiler in Go.

Peter Steinberger 🦞

@steipete

X

Kind of amazing that it took Microsoft a team and two years to rewrite the TS compiler in go, and now you can just let an agent rip and get sth significantly better within weeks.

The business read is that the scarce resource here is not agent capability, it is supervision design. One founder, two agents in an adversarial pairing, and a check-in cadence measured in days. These are Steinberger's own accounts of his own projects, not benchmarked results, but the pattern is cheap enough to copy.