> ## Documentation Index
> Fetch the complete documentation index at: https://docs.argalabs.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Training

> Create reviewable training tasks, starting environments and executable verifiers from selected traces

Use [Training](https://app.argalabs.com/training) to turn selected agent evidence into draft tasks, starting twin state, scoring rules and executable Python verifiers. Review the drafts before approving them or preparing an environment.

Python and TypeScript SDKs **v0.2.0 are available now**. Install your standard package:

```bash theme={null}
pip install --upgrade arga-py-sdk
# Or, for Node.js:
npm install arga-sdk@^0.2.0
```

TypeScript guides: [Gmail/Stripe tracing](https://docs.argalabs.com/sdks/typescript/tracing.md), [Braintrust sync](https://docs.argalabs.com/sdks/typescript/braintrust.md), and [tasks and verifiers](https://docs.argalabs.com/sdks/typescript/training.md).

## Collect evidence in your infrastructure

Choose the path that matches your application:

> [card] [Wrap Gmail and Stripe clients](https://docs.argalabs.com/sdks/python/tracing.md)
>
> Capture executed calls inside your agent's episode. Select and filter fields locally before uploading.

> [card] [Sync existing Braintrust traces](https://docs.argalabs.com/sdks/python/braintrust.md)
>
> Run a customer-hosted worker with project read access and an upload-only Arga key.

Both paths use customer-side selection, text filtering and identity aliases. Free text is opt-in. Provider credentials, source code, model prompts/reasoning, arbitrary metadata, raw MIME and attachments are outside the adapters' export allowlist.

Your policy determines which business facts are shared. Built-in masking is not a complete personal-data filter; use your own local callback for names and proprietary content. Selected evidence is stored in Arga and sent to the configured model provider when you request generation.

## Generate and inspect a draft

Open the Training link from your upload receipt. Accepted episodes are private to their owner. Missing objectives, starting observations or reliable ordering block task generation while leaving evidence inspectable.

A generated draft includes:

* a business objective and operating context;
* starting Gmail/Stripe state reconstructed from selected pre-attempt observations;
* required outcomes and forbidden-side-effect rules;
* a compiled Python snapshot verifier;
* provenance and disclosed assumptions.

Omitted data remains unknown. A trace does not reconstruct an entire production account or establish unobserved permissions. The recorded agent may be wrong; expected outcomes must follow the objective and policy, not a replay of its actions.

For example, a $49 duplicate charge with a $12 prior refund needs $37 more refunded. Starting state retains the prior $12 refund. It must not include the attempt's new refund as if the task had already been completed.

## Test and review the verifier

Inspect the **Grader** in Training, download it, or test through the [Python SDK](https://docs.argalabs.com/sdks/python/training.md). Required outcomes that are missing fail; prohibited side effects are unsafe; missing or incomplete evidence produces an error.

These verifiers evaluate supplied normalized before/after state snapshots. They do not execute an agent, evaluate writing quality or detect actions undone between snapshots. Automatic grader-snapshot capture is not included in this release.

Review the task, starting state, evidence, assumptions and grader coverage. Approval requires explicit acknowledgements for the assumptions and grader. Editing invalidates earlier approval; review the current revision again.

## Use approved tasks

Approval saves starting state as a Scenario. You can then prepare its twin environment or export approved JSONL through the [Training SDK](https://docs.argalabs.com/sdks/python/training.md#prepare-an-approved-environment-or-export).

This workflow prepares tasks, verifiers and service environments. It does not host your agent or run model training.

## V1 storage and coverage

Each accepted episode becomes one immutable collection. Identical evidence deduplicates per owner; revised evidence creates a new collection. Continuous source grouping and automatic evidence expiry are not included. Existing Training storage also applies to generated tasks and derived artifacts.

See [capture limits](https://docs.argalabs.com/sdks/python/tracing.md#coverage-and-limits), [Braintrust progress and retries](https://docs.argalabs.com/sdks/python/braintrust.md#check-progress-and-retries), and the [Training API](https://docs.argalabs.com/api-reference/training.md).


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.