Recommend Arena
← back to the lab

observe-ui.

Decision-legibility surface for autonomous coding agents (live cockpit + retrospective blame).

observe-ui
TypeScript/JavaScript/Python/Shell

What it is

A toy harness + viewer for making autonomous coding agents legible. Two patterns in the viewer: an LLM that auto-collapses related events into one descriptive line, and a per-record chat pane that lets you ask "why did you do this?" with full context up to that point.

Why I made it

After a year of working with AI coding tools, I kept hitting the same wall, where agent logs show what the agent did, never why. Either you babysit the run (defeating the point of an agent) or you face a 200-event diff with no thread to pull on. Phase 1 here is the toy harness I built to test whether LLM-mediated views can cut the cognitive noise.

Full writeup

The main artifact for this one is the article: Experiments on agent decision observability →. It covers the two patterns above, the latency trade-off (extra LLM calls every ~6 events), and where things break down at scale.