multirun

Point it at a source (a repo at a commit), give it a task, and it fans out N isolated runs of a coding agent — each run gets its own git worktree and its own Neon database branch — then captures what each run did and lets you compare them side by side.

The db-branch-per-run is the point: all N runs start from byte-identical database state, so diffing each branch back against the parent tells you what the agent actually did to the data, instead of reading N transcripts.

run = (source, task, variant) -> artifacts

Agents run through Anthropic's Agent SDK on subscription auth — there is no api-key path anywhere in the runner.

layout

core/            the python package (harness/), its docs and tests
  harness/
    cli.py           submit / status / logs / teardown, and `local`
    config/          schema + global/per-run merge
    orchestrator/    plan -> spawn (k8s job per run) -> reap
    runner/          the deliberately dumb part: config in, artifacts out
    resources/       acquire()/release(): neon branch, worktree, storage
    diff/            branch vs parent: schema and rows
    results/         run history, on Neon, queryable in SQL
tasks/           content, not code: one dir per task (prompt, source, seed, checks)
dashboard/       the read-only web ui: results db + artifacts volume, htmx, no build step
infrastructure/  images, chart, argocd — written; deploying is its own phase

start here

core/docs/design.md for how it thinks, core/docs/runbook.md to run the smoke loop on your machine, core/docs/config-reference.md for the yaml.

python -m venv .venv && .venv/bin/pip install -e "core[dev,results]"
.venv/bin/multirun local core/configs/smoke.yaml