Ur is a framework for hill-climbing a model with many agent sessions. It's designed for a kind of 'dark factory' approach to model development, where agents autonomously initialize research.
orchestrator ─► n workers ─► held-out benchmark ─► orchestrator combines the best ─► n workers ─► ...
Nothing written in prose passes from one worker agent to another, to contain memetics that might keep a configuration in a local minimum. Every worker agent starts in a fresh session with a copy of the current best model and a short brief.
The design follows two findings from Fork-and-Flush:
- Autoresearch agents settle into "idea basins".
- Starting fresh sessions on top of the best workspace gets them out. Summarizing their history doesn't.
git clone https://github.com/abiome-org/ur && ./ur/install.shinstall.sh installs the ur command (with uv, pipx or pip) and runs ur install, which installs the skill. The skill is copied to ~/.config/agents/skills/ur and linked from every other agent skills directory it finds under ~/.*/ and ~/.config/*/. Ur needs Python 3.11+ and git, and has no other dependencies.
In a project, tell your agent: "I'd like to initialize abiome-org/ur for this project."
.ur/
config.toml objective, splits, scorer, training command, budgets, agents, devices
brief.md what every worker is told; frozen after the first round
eval/ the scorer, sealed at setup
checks/ directories that check.expect scores (trivial predictors, reference answers)
state/ campaign state (gitignored)
The heavy files (code snapshots, runs, checkpoints, workspaces) live in a store outside the project, at ~/.local/share/ur/<project>-<hash> by default. Code snapshots go in a bare git repo there, checkpoints in per-run directories. Each workspace is a plain copy with no history, so siblings can't reward-hack by peeking at each others work.
ur status [--json] campaign phase and the next step; in a workspace, your time left and scores
ur init write a config template for this project
ur check [--probe] setup checks, then score the cheat probe (--accept: its gain is legitimate)
ur brief [--draft] print the brief, or a draft built from the config
ur round open a round: one workspace per worker, copied from the base
ur prompt <workspace> print a workspace's prompt, to pass unchanged to a fresh session
ur spawn <workspace> start a fresh session in a workspace (agents.spawn = "command")
ur wait [--timeout S] block until the current phase has its next step ready
ur close end the round and score the workers on held-out
ur select score the candidates and set the new base
ur run train and score this workspace under the per-run budget (dev split)
ur commit [--candidate] keep the last run
ur reset [<node>] put the workspace back to your last commit, or to a node
ur diff [<a>] [<b>] code changes between nodes, or between a node and this workspace
ur apply <node> three-way merge a node's changes into this workspace
ur log runs in this workspace, or the campaign's base history
ur export <node> <dir> copy a node's code and checkpoints out
ur archive end this campaign and keep its records
ur install [--uninstall] install the skill for the agents on this machine
uv run --with pytest pytestThe tests drive whole campaigns on examples/toy, a tiny regression problem in pure Python.