Some updates on OpenClaw and agent orchestration. I’ve tried all of this extensively, so these are notes from use rather than from the brochure.
Lossless Claw: good. It’s a plugin that replaces OpenClaw’s built-in sliding-window compaction with a DAG-based summarization system that preserves every message while keeping the working context inside the model’s limit. Every message goes into SQLite, older stretches get summarized, the summaries roll up into higher-level nodes, and the agent can drill back down to the original text when it needs to. Being able to retrieve older messages helps in long agent sessions.
The default RAG setting in OpenClaw: also good. With an embedding provider configured, its memory search runs hybrid retrieval — vector similarity for meaning, keyword matching for exact terms like IDs and code symbols. Leave it on.
Paperclip.ing and Gas Town: not ready for prime time. Both are ambitious in the right direction. Paperclip calls itself “the open-source app everyone uses to manage agents at work” — an org chart for agents, with roles, titles, reporting lines, permissions, and a monthly budget per agent that stops it cold when the money runs out. Gas Town is Steve Yegge’s orchestration system for Claude Code and other agents, built on the premise that four to ten agents become chaotic and that it can scale comfortably to twenty or thirty. The ideas are sound and the software isn’t there yet, at least for the way I work.
Stitch from Google is also very interesting for UI/UX. One caution: you may need to start it up fresh a couple of times, because it loses its bearings after just a couple of iterations. Treat each session as a sketch rather than a project.
Oh — and the GSD and AutoResearch code repositories are fantastic templates. GSD gives Claude Code a discipline of phases, each run in a fresh context so quality doesn’t rot as the window fills. AutoResearch is the loop that changes one thing, measures, keeps or discards, and goes again. You can adapt them for anything, and they’re great.
