Arbor — Autonomous Optimization via Hypothesis Tree Refinement

Overview

This skill runs an Autonomous Optimization (AO) loop: starting from an existing artifact and a measurable objective, improve it through many rounds of experiment and evaluation — without step-by-step human supervision and without overfitting to the feedback signal. It's the right tool when the bottleneck isn't writing one good change, but organizing dozens of trials so that lessons accumulate instead of evaporating.

It implements Hypothesis Tree Refinement (HTR) from Arbor (Jin et al., 2026). The key idea: keep the research state in a persistent hypothesis tree rather than in conversation history. Each node binds a hypothesis, the distilled insight it produced, and a pointer to the artifact version that realizes it. You play the long-lived coordinator that owns this tree and decides where to search; short-lived executor subagents test one hypothesis each in isolated git worktrees and report back. A held-out merge gate admits a change only when it improves on a test evaluator the search never optimized against. This is what turns trial-and-error into cumulative, auditable research.

Use the scripts/tree.py state manager for all the bookkeeping (creating nodes, writing evidence, propagating insights, pruning, the merge gate, the Observe projection). It keeps the state consistent and frees you to spend judgment on what the evidence means.

When to use this skill

Reach for Arbor when the task is iterative improvement of a concrete artifact under an evaluator:

• Model training: optimizer/architecture/recipe changes to lower loss or hit a target in fewer steps.
• Harness/agent engineering: raising pass rate or accuracy of an agent loop, search harness, or tool-use scaffold.
• Data synthesis: improving a generation/filtering pipeline judged by downstream model behavior.
• Benchmark optimization: MLE-bench / Kaggle-style "improve the submission" tasks.
• Prompt/system optimization where you can score outputs automatically.

The distinguishing signals: there's an artifact you can modify, an objective, a way to score candidates, and you expect to run many experiments. If the user only wants a single fix or a one-shot answer, this is overkill — just do the work directly. If they want open-ended ideation with no evaluator, use hypothesis-generation or scientific-brainstorming instead.

The AO setup — pin this down first

Before any experiments, establish the task tuple (M_0, O, E_dev, E_test). Getting this right matters more than any later decision, so confirm it explicitly:

• M_0 — initial material: the artifact to improve (a repo, a script, a config, a prompt). Make sure it's under git and currently runs.
• O — objective: the natural-language goal and the metric direction (maximize accuracy? minimize loss/steps?).
• E_dev — development evaluator: a command you can run freely during search to score a candidate. Fast, repeatable.
• E_test — held-out test evaluator: a separate evaluator (different seeds, different split, or a larger run) used only at the merge gate. It must not be used as a search oracle — that's the whole point.

If the user hasn't given you a clean dev/test split, construct one and say so. The dev/test separation is the mechanism that catches overfitting: a candidate that wins on dev but not on test isn't a success, it's a warning that you're exploiting the feedback signal. Without it, autonomous search reliably overfits.

Initialize the run:

python scripts/tree.py init \
  --objective "Improve BrowseComp answer accuracy on the search harness" \
  --dev-eval "python eval.py --split dev --n 50" \
  --test-eval "python eval.py --split test --n 300" \
  --material "." --metric-direction max --branching 3 --max-depth 2 --budget 12

--branching is how many sibling hypotheses you propose per parent; --max-depth 2 keeps directions at depth 1 and concrete interventions at depth 2 (the paper's default); --budget is the number of coordinator cycles. Start small (10–20 cycles) — structured search beats brute force, and you can extend if progress is still being made.

The coordinator loop

You run repeated cycles of six steps. This is the heart of HTR; do not collapse it into ad-hoc editing. Run python scripts/tree.py cycle once per cycle to track the budget.

1. Observe

Begin every cycle by re-grounding in the tree, not in your memory of the conversation:

python scripts/tree.py observe

This prints the objective, global insights, the active frontier (selectable hypotheses), executed nodes with their evidence, pruned lessons (negative constraints), and the current best artifact. Treating the tree as the source of truth is what keeps you coherent over a long run, after context compression has thrown away the details.

2. Ideate

Pick a promising parent and propose a few child hypotheses under it. Condition on the tree's evidence — this is the difference between Arbor and random search:

• Validated insights are assumptions you can build on.
• Pruned nodes are dead ends to avoid.
• A "half-right" result is a starting point for a sharper hypothesis, not a reason to abandon the direction.

Each hypothesis should be a falsifiable claim about how changing the artifact will move the metric, not a vague intention. Depth-1 nodes are broad directions ("the search harness loses correct answers it already retrieved"); depth-2 nodes are concrete, executable interventions ("run K=5 independent rollouts and aggregate by evidence dossier instead of majority vote").

python scripts/tree.py add-node --parent n0 --hypothesis "Verification, not retrieval, is the bottleneck: candidates are found but discarded"
python scripts/tree.py add-node --parent n4 --hypothesis "Decompose the question into atomic constraints and verify each independently"

3. Select

Choose which pending leaves to run next. Selection is not pure score-maximization — pick a hypothesis because it has strong prior evidence, because it would resolve an ambiguity its siblings exposed, or because its failure would clarify an important assumption. Frontier control under delayed feedback rewards informative experiments, not just promising ones.

4. Dispatch

Run each selected hypothesis as an executor subagent in an isolated worktree (use the Agent tool with isolation: "worktree", or have the executor create one with git worktree add). Isolation matters: parallel experiments must not clobber each other or the current best, and exploratory changes stay quarantined until they pass the merge gate.

Dispatch siblings in parallel (multiple Agent calls in one message) when they're independent — comparative evidence within one direction is exactly what makes later pruning and abstraction possible.

Give each executor a tight, hypothesis-bound brief. See references/executor-brief.md for the full template. The contract that makes HTR work: the executor may not change the hypothesis when the metric stalls. It repairs its own code and reruns, but h_n is fixed — otherwise the returned score is no longer evidence about the assigned node and the tree's semantics break. The executor returns exactly four things:

• dev_score — the dev evaluator result (for selection);
• result — a factual summary of what happened;
• insight — the distilled, reusable lesson (why the result supports, weakens, or bounds the hypothesis);
• branch_ref — the git branch/commit/worktree path holding the artifact.

Mark a node running before dispatch (tree.py set-status --node n5 --status running) so the Observe projection stays accurate.

5. Backpropagate

When an executor returns, write its report into the node, then abstract the lesson upward:

python scripts/tree.py set-evidence --node n5 --dev-score 70.0 \
  --result "K=5 dossier aggregation reco