Skip to Content

ablate try

Smallest real run: 3 tasks × 3 trials × 2 arms on the cheap model. ≤ $2, ~6 min.

ablate try ablate try CLAUDE.md --yes

try is a thin wrapper around compare with --trials 3, --cheap, and --max-spend 2 fixed.

Arguments & flags

Argument/flagDefaultWhat
CONFIG_Aauto-detectedSame auto-detection as init.
--tasks DIR.ablate/tasksDirectory of task YAML files.
--tasks-only IDSnoneRun only these task ids or globs.
--tasks-tag TAGSnoneRun only tasks carrying any of these tags.
--concurrency N3Parallel trials.
--yesoffSkip the confirm prompt.

What it does

Requires the claude CLI on PATH, authenticated or with ANTHROPIC_API_KEY set — it calls a real agent inside an isolated git worktree.

It is deliberately underpowered, and deliberately honest about it. Three tasks cannot produce a confidence interval at all (eight minimum), so every verdict will read INCONCLUSIVE. That’s the correct answer at this size, not a failure — it proves the plumbing works against a real agent without claiming to measure anything. A tool whose selling point is statistical conservatism cannot open with a demo that quietly produces a believable number.

Last updated on