Skip to main content
Use a full Arga credential with the training resource in v0.2.0+. Upload-only trace keys cannot read collections, generate tasks, approve them or prepare environments. Begin with SDK evidence or Braintrust sync.

Generate a draft

Use the collection UUID from an upload receipt or Training URL. Generation can take up to several minutes. The draft includes the task, starting state, evidence, assumptions, scoring rules and a compiled Python snapshot verifier. Incomplete coverage and missing objectives or starting state block generation; omitted state remains unknown. Expected outcomes follow the objective and operating policy. For a $49 duplicate payment with a prior $12 refund, the task should preserve the original payment, start with the prior refund, and require $49 total refunded. It should exclude the recorded attempt's newly created $37 refund from starting state.

Inspect and test the verifier

The remaining examples share client, dataset and task:
The verifier is compiled from the proposed rules, rather than model-written executable code. It supports existence, absence, equality, substring, count and unchanged-state predicates. Generated verifiers remain Python artifacts even when called from TypeScript. Capture normalized before/after snapshots containing every collection, record and field the rules require. Each collection is { "complete": true, "records": [...] }, nested under provider and resource names. Keep provider keys such as payment_intents in snake_case and use task-consistent IDs. See the shared snapshot contract.
You can also run the downloaded verifier locally:
Snapshots are limited to 2 MB, 5,000 records per collection and 30 rules. Declare empty collections explicitly. The SDK does not capture these snapshots automatically; the example in a grader bundle is illustrative. Snapshot verifiers do not judge writing quality, run your agent or detect transient actions undone between captures.

Edit and rebuild

Submit the full specification and keep its objective, starting state, provenance and rules consistent. Edits invalidate earlier approval. Use the returned revision for later tests and review; stale revisions return HTTP 409.

Review and approve

Open the collection in Training. Review its task, starting state, scoring, source evidence, assumptions and compiled grader. After that review:
Do not automate the acknowledgements without inspecting the actual revision. Approval saves a reusable Scenario. Rejected or unreviewed tasks cannot be exported as approved tasks or used to prepare environments.
If you edit or rebuild a task, review its current revision again before preparing an environment. These methods prepare data and service environments; they do not run customer agents or train a model. Collections remain private to their owner within the existing Training storage lifecycle. Requested generation sends selected evidence to Arga’s configured model provider; v1 adds no automatic expiry.