An AI optimizer now claims a 2.5x performance edge over Claude Code and Codex, and that gap matters more than a leaderboard screenshot suggests. By wiring its decisions through Arbor, the system treats optimization not as a monolithic black box but as a series of discrete bets whose impact can be audited.
The core shift is architectural, not cosmetic. Arbor separates strategy from execution by assigning each optimization attempt to an isolated git worktree, so every change set lives in its own branch of reality while the main repository stays untouched. That structure turns what used to be a blur of edits from large language models into labeled experiments, tied to specific hypotheses about latency reduction, memory footprint, or algorithmic complexity.
This traceability rewrites how engineering leaders think about AI assistance. Instead of treating generated patches as one-off suggestions, teams can run a closed-loop cycle: propose a strategy, let the optimizer generate and execute code in a sandboxed worktree, then measure outcomes against performance baselines and regression tests. Git internals, from object storage to tree references, become the spine of an attribution system that links every micro-optimization to metrics that either move or do not move the needle.
The commercial implication is blunt. A 2.5x edge is not just about faster code generation; it concentrates the learning rate of the entire organization in a single optimization engine that remembers which strategies failed, and why, encoded in commit history rather than in slide decks.
