Autoresearch with hi

Let coding agents try many ideas while hi checks and scores every attempt itself and keeps only what works. One round to pick the best of n, or rounds overnight that commit only gains.

Agents vary a lot from one attempt to the next. Some attempts are right, some are wrong, and some look right without being right. hi agent best-of turns that variation into progress: several agents try the same task in separate boxes, hi checks and scores what each one did, and only the results that pass and score well are kept.

It works in two ways:

	Best of n	Rounds
Use it for	one task, a few tries, pick the best	a number to improve over many tries: a loss, a benchmark, a Sharpe ratio
Judged by	your check (`make check`), then a ranking	a score command that prints a number
Who keeps a result	you, with `keep`	hi, only when the score beats the best so far
How long	one round, minutes	rounds until a limit, often overnight
Command	`hi agent best-of 3 "<task>"`	`hi agent best-of 3 --rounds 50 --score "make score" --lower "<task>"`

hi agent best-of 3 --agents claude,codex "make the backtest loader 2x faster"
hi agent best-of 4 --rounds 100 --score "make score" --lower --edit train.py program.md

Rounds are Karpathy’s autoresearch loop (try an idea, measure it, keep it only if it’s better, repeat), with the judging moved from the agent to hi.

Why you can leave it running

hi judges, not the agent. When an agent ends, hi itself runs your check and your score in its box, on the files as the agent left them. An agent that says its tests pass, or that it improved the number, is checked, not believed.
Only hi commits. In rounds, agents work on scratch branches. hi adds a commit to the run’s own branch, best-of/<run>/best, only when a result beats the best so far. Nothing is merged into your branch or pushed; git merge is your step.
Boxes contain the work. Every attempt runs in its own rootless container with only the project, a network limited to what you allow, and no credentials. --edit limits which files a result may change, and hi discards results that touch anything else, such as the evaluation.
Limits are yours. You confirm the cost before anything starts, and a run ends at --rounds, --for, --budget, or --patience, or when you stop it.

One round, step by step

hi starts n boxes from the same commit (for rounds: the best so far), each on its own branch, with --agents taken in turn.
Each agent works on the task, unattended, without permission prompts.
When an agent ends, hi stops whatever it left running, snapshots its files, then runs --check and --score on that snapshot.
hi ranks the boxes. In best of n, you pick one with keep. In rounds, the best score is committed if it beats the best so far, and every box is removed.
In rounds, the next round’s task tells the agents the best score and what every earlier attempt tried and scored.

Where to start

Several attempts at once: best of n, the table, and keep.
Rounds that keep only gains: scores, limits, watch, stop, and overfitting.
Start from a template: hi init autoresearch-ml, a project already shaped for rounds, with a tested evaluation and a holdout.
Simple examples and advanced examples.

Every attempt is a full agent run, so a run costs about n times the tokens of one, times the rounds. hi shows the plan and asks before it starts; agents pass --yes only after their person agreed.