Sidecat Sidecat, restarted
Sidecat is on its third build of the same idea. This is a short note on why the current generation strips back instead of adding on.
The same problem, three times
The first generation was a Python prototype. It proved the real shape of the problem: AI-assisted work that matters cannot just live in a chat log — it needs durable records, resumable orchestration, and governed external effects.
The second generation made that usable as a Go runtime, with work records, gates, evidence, and receipts. It worked, and it taught us the most expensive lesson: a sprawling surface is its own kind of debt. Every extra command and concept is something a person — or another agent — has to read past to do the actual work.
One loop: queue, run, review
So the third generation is deliberately smaller. The shape that survived is a single loop:
- Queue the work as a durable, governed record — what it is, what it targets, and the conditions that must hold.
- Run it through explicit authority, keeping what happened next to what the agent was shown, and leaving receipts for anything that spends trust.
- Review it from that one record — turning it into review, replay, and a resume brief — so the work can be checked and picked up later without transcript archaeology.
That is the whole shape. The durable record is the product; everything else is a lens over it.
What we are not claiming yet
This is early R&D, in the open. Sidecat is a working direction, not a packaged release or a hosted product, and the fuller product vision is still being written. We are building it by using it — and writing down what holds.
— The Sidecat Dev project