How an autonomous agent searches for the groundrule — and what we think makes the loop different.
We don't write the rules.
We just discover them. With AI.
Finding the best groundrule for Le Petit Renard means searching a vast space of possible rules. We hand that search to an autonomous agent: it proposes, tests, and keeps what works — on its own. Below: the loop, and the ideas that make it go further.
We don't hand-write the groundrule. An agent runs an autoresearch loop: read what's been tried, propose a rule, test it on every night, score it, keep it only if it beats the best — then go again, on its own. It's the agentic setup Andrej Karpathy points to as where AI is heading.
Two roles, split clean. The agent owns the loop and runs it autonomously. We supervise from outside: set the levers — patience, model, library — curate the inspiration, and decide when to pivot or start a fresh run.
Each round keeps only what beats the best so far, and builds the next groundrule on it.
A search you can't see is one you can't steer. Three live views show what the agent is doing — so you know when to let a run breathe and when to pivot, swap models, or change the library.
The control room. Every candidate the agent proposes lands here the moment it's scored — newest on top — beside its total wait and whether it beat the best so far. The chart tells the same story: scores falling round by round, the running best stepping toward the target.
It's where you watch a run breathe — and catch the moment it stalls.
Open the dashboard →One service night, every candidate's schedule laid side by side — then stacked into a single overlay. Where the agent keeps landing on the same move, the bars pile into a column and the cluster jumps out; where it's still casting about, they scatter.
A read on what the search has settled on, at a glance.
Open the grid →No candidate appears from nowhere. Each one builds on a parent — and the lineage draws the whole search as a family tree: the lines that refined step by step, the pivots that struck out in a new direction, and the single path that reached the champion.
The search's entire history, in one picture.
Open the lineage →Give the agent a patience — how many rounds it may go without a new best. Spend it, and it must pivot: drop the current line for a genuinely different groundrule. Our main guard against rabbit-holes.
A search isn't one fixed run. Each new run inherits the champion — the best groundrule so far — and climbs from there, while you change the model, the patience, the library around it.
The library is an optional, human-written input — notes, framings, even a personality handed to the agent before it thinks. It doesn't fix the answer; it shapes the frame the search starts from.
The loop talks to any OpenAI-compatible endpoint, so the proposing model can run locally — LM Studio, or anything serving the same API. Point it at your own hardware and the search runs unmetered: no rate limits, no quota walls.
Each candidate faces a battery of service nights, every one leaning on a different station — so a winning groundrule has to be good all round, not just on the easy nights.
Three more nights stay sealed. The kitchen never trains on them, so they're the exam that proves a groundrule generalised rather than memorised.
Start on the research dashboard to watch the next groundrule get found, or revisit the problem it's solving.