AI optimizer beats Claude Code, Codex by 2
Researchers introduce Arbor, a framework that structures AI optimization as a cumulative learning tree instead of isolated trial-and-error loops, delivering 2.5x verifiable gains over Claude Code and Codex on the same compute. It solves the core AO bottleneck: agents forgetting what they’ve learned across long-running experiments.
Transcript
Justy Okay, this one’s wild. Arbor beat Claude Code and Codex by two and a half times on the same budget, just by fixing how the agent remembers what it’s tried.
Cody Yeah… that’s the kind of claim that usually means they found a trick, not a breakthrough.
Justy I know, I know. But the trick here is basically ‘stop forgetting your own work’—
Cody That’s it? That’s the breakthrough?
Justy No, the breakthrough is the tree. They turned the optimization loop into an actual tree that stores hypotheses, experiments, outcomes—so the agent doesn’t just scrollback and forget everything after context window.
Cody Right. That’s… actually the real problem.
Justy How was your week, by the way? I feel like I haven’t seen you since that disaster in DC
Cody I got back late last night. Some flight delays. My back still hates me… anyway —
Cody So Arbor keeps the state, lets it prune bad paths, pursue the ones that actually worked.
Justy Exactly. And the numbers back it up—two point five times on real tasks, same compute.
Cody Fine. But does it scale? A tree’s great until you’ve got a thousand branches and the agent’s just managing the tree instead of optimizing.
Justy That’s your classic Cody move. It works, so you assume it’s about to break.
Cody I mean, it’s a fair question. If the overhead of maintaining the tree eats the gains, then it’s a wash.
Justy Maybe. Or maybe the tree’s the point—the thing that turns AO from a noisy loop into actual research.
Cody Okay, okay. And the repo’s just called Arbor, no cute branding?
Justy Yep. Paper’s from Renmin University and Microsoft Research. Jiajie Jin’s quoted in there saying a loop isn’t the same as progress.
Cody That tracks. If the metric’s hackable, more time just gets you more garbage faster.
Justy So Arbor’s the harness that stops the garbage. And if you’re running AO in production and drowning in trial-and-error noise…
Cody This is the harness thing, not the model thing.
Justy Exactly. That product-is-the-harness take we had last month.
Cody Hm. So if your agent’s failing, it’s probably the loop design, not the brain.
Justy And now there’s a tree for the loop. Anyway —
Justy I’m just saying, if this works, it changes how teams actually ship AO.
Cody Sure. And if it doesn’t, we’ll all forget about it by next week.
Justy Cody, continuity has ruined your brand.
Cody I miss contradicting myself freely.
Justy Love you too. Catch you next one.