Interactive Autoresearch
Agents already execute well. That was never the hard part. Sapling pairs that execution with your taste — proposing what to run and why, grounded in what you've already tried, and only moving once you say go.
runs on whatever's actually worth running
you › what should we try next on the ablation?
✻ last 3 runs plateaued around 61% held-out once we dropped the aux head. checking whether that's a capacity or a schedule problem before proposing anything new…
worth isolating: rerun the aux-head config with a 2x longer schedule before we conclude it's capacity. Same seed, same split.
sapling wants to run: launch_experiment
config=aux_head_v3.yaml epochs=200 (was 100) budget≈$4.10
[y] once [s] session [p] this project [a] every project [n] reject
● log_write(node=run-042, dev_score=0.643, held_out_score=0.611)
✓ logged run-042 · dev 0.643 / held-out 0.611 (schedule, not capacity, aux head kept)
Sapling inherits its execution from autoresearch systems and its taste from coding agents, one lineage per half of the problem.
execution
taste-aware
both
Agents are already excellent at running things. Sapling is built around the part that's still yours: the taste for knowing what's actually worth trying.
Every proposal is checked against what you've already tried and what the real literature supports. Not a generic best-practice guess: a claim tied to evidence.
Cadence, from checking in on every step to running with minimal interruption, is yours to set live, mid-session. Sapling never decides on its own how much rein to take.
Every outcome is logged locally, dev and held-out, tied to what backs it up. A result can't be marked "promoted" without both. It's enforced automatically.
Pair mode: the default and primary way to run sapling. Always in this order — propose before approve, approve before run.
Sapling reads your research log and proposes a next step, with its reasoning shown inline, not hidden behind a spinner.
You see the exact command and its cost before anything runs. Approve once, for the session, for this project, or for every project.
Sapling kicks off the experiment and reads results back from whichever tracker you already use: W&B or MLflow, read-only.
The outcome, dev and held-out, is written to a local, append-only research log. Nothing overwritten, nothing silently dropped.
Free to start, no card required. Upgrade whenever you want the full model list.
Create your account →