Arbor
Arbor is an autonomous research agent that is given a benchmark and a goal, then proposes hypotheses, edits code, runs real experiments, and keeps only the changes that hold up on held-out data, building a tree of hypotheses instead of forgetting what failed.
Arbor is a generalist autonomous research agent. Given a benchmark and a goal, it proposes hypotheses, edits code, runs real experiments, and keeps only the gains that survive held-out data, growing a hypothesis tree that preserves failed attempts instead of discarding them.
The project reports results across six tasks, including BrowseComp, Terminal-Bench 2.0 and a math-reasoning benchmark, and describes Arbor as recalling concrete findings from its own earlier runs at the start of a similar task. It can ground research in prior work through a public literature-search API, check the novelty of an idea before building it, and ships a native CLI runtime plus an agent skill suite for Codex and Claude Code.