Research a draft task by task

Inspect and improve each step before freezing and running the complete pipeline.

A registered draft can run one task at a time in a private research workspace. This is useful when you want to inspect an extraction before sending it to a cluster, tune the extraction, and then let the cluster consume only the newest successful result.

Run one task

group = workspace.workbench_group(
    "event-recommendations", "v1", "extract-events", tasks=["extract-events"],
)
group.run_sample(
    # Omit sources when these rows already exist in the workspace Space.
    sources={"headlines": sample_headlines},
    sample_size=100,
    sample_seed=7,
    max_cost_usd=2.00,
)
extracted = group.rows()

The browser runs and stops single tasks on every draft version. From Python, a Workbench group of one task runs that task's sample, and group.rows() returns the rows the latest sample produced. Sample size counts complete groups for Union and grouped Apply; other tasks sample rows per input Space. Groups include every member in the current input, which may itself be an upstream sample. Receipts report sampled_groups when applicable and actual rows in input_counts. The seed makes the choice reproducible.

Current results and reruns

Only a successful result becomes current. Rerunning a task successfully replaces the result shown to downstream research. A failed or stopped run leaves the preceding successful result intact. Research history retains cost, duration, status, and row counts.

Improve the draft from evidence

After inspecting a task result, edit the draft in the browser, register its edited declaration with workspace.register_version(spec), or edit exported public code. Researchers and coding agents can refine a prompt, exact transformation, normal grouping rule, semantic gate, or task topology. Read the Planner rationale for expected hand sizes, costs, uncertainty, and known failure modes before retaining a preliminary rule. Validate any changed cutoff on a separate hand; a visible sample is evidence to investigate, not proof that a rule is safe.

Automatic second-, third-, or later-shot DAG improvement is experimental rather than a first-class Planner promise. The separate optimization surface remains available after execution and review: it automatically tunes supported runtime parameters and proposes external prompt or parameter changes for explicit admission. Neither research execution nor optimization advice silently mutates a draft.

Dependency order

A downstream task unlocks only after every task that produces its inputs has a current result. Barriers are also respected. If you rerun an upstream task, dependent results become stale and must be rerun before later tasks can proceed. You cannot override an interior Space with uploaded rows; rerun its producer instead.

Research results are isolated

Research rows do not appear as production session output and are never written to connected sinks. The interface shows the version's ordinary logical Space names; experimental storage is not part of your pipeline definition.

Declare when satisfied

Declaration freezes the draft and its research history, then discards experimental rows, logs, relationship evidence, and unused draft prompts. It does not promote experimental data into production. A declared run starts from its bound sources and final prompt bindings. Compatible model-call cache entries remain reusable inside this workspace; sibling workspaces and organizations are fully isolated. Before declaration, researchers may inspect current research output without turning it into pipeline input.