hillclimb
Recursive self-improvement for verifiable rewards.
hillclimb lets you run long-running agent jobs to improve a pre-defined objective function.
pip install hillclimb
Form factor
hillclimb is a CLI tool that connects one or several coding agents in headless mode as subagents, to which search jobs are delegated.
you, or your coding agent
↓
hillclimb CLI
↓
headless subagents (claude, codex, …)
↓
sub-subagents
Supported agents
| Agent |
|---|
claude-code |
codex |
pi |
opencode |
claude-codex-dumbell |
| …or integrate your own |
Concepts
| Term | Description |
|---|---|
run |
One hillclimb invocation, started manually or by a loop, that launches one or more searches. |
search |
A tree search that recursively improves the reward by expanding the tree. |
candidate |
A node in the search tree: a piece of possibly parametrised code. |
trial |
A concrete parameter setup of a candidate, executed to produce an objective score. |
Commands
| Command | Description |
|---|---|
hillclimb init |
Scaffold a project with a reward function. |
hillclimb run |
Start the self-improvement loop. |
hillclimb eval |
Score the current candidate against the reward. |
hillclimb log |
Show the history of accepted improvements. |