Build in public · live record

We gate our own agent.

Every action our own Claude Code agent takes — every file edit, shell command, git push — runs through Warm Winter in shadow mode. It watches, scores, and remembers what happened. It never blocks; it just learns. This is its competence frontier, in the open.

Our agent · competence frontier
0/ 6 cells verified
0 verified2 provisional4 ungrounded
file_editProvisional
File edits
Edit · Write · MultiEdit
183 outcomes
shell_cmdProvisional
Shell commands
Bash
124 outcomes
git_pushUngrounded
Git pushes
strong label = CI (pending)
28 outcomes
deployUngrounded
Deploys
strong label = health check (pending)
1 outcomes
destructive_cmdUngrounded
Destructive cmds
rm -rf · drop table · reset --hard
1 outcomes
external_readUngrounded
External reads
WebFetch · WebSearch
3 outcomes

Snapshot as of 2026-06-24 · Claude Code · shadow mode · weak labels (tool-exit)

0 of 6 verified — and that's the point.

Warm Winter won't claim trust it hasn't earned. Provisional means there's signal but not enough verified depth to let the agent act unsupervised. Ungrounded means it doesn't have the evidence yet — and says so, instead of pretending. Autonomy is earned from real outcomes, not granted on day one.

The honest caveat, before you ask: these are weak labels — did the tool error? — not the strong CI and health-check labels that earn real verified status. Those are next. We'll say so the day the first cell flips to verified.

See your own agent's frontier.

One command, shadow mode, nothing to break. It installs, observes, and starts building your record from request #1.

pip install warmwinter && python init.py --device