Async vs sync workflow

Async vs sync workflow in Stratix — when work is queued, when it's immediate, what to expect.

Stratix has two operational modes: synchronous (immediate response) and asynchronous (queued, results stream in). Knowing which mode applies prevents most "why is this taking so long?" surprises.

Sync operations

These return immediately:

• Browsing the catalog

• Fetching a model, benchmark, or saved evaluation

• Reading a trace

• Creating a judge or scorer (definition only)

• Generating an API key

Async operations

These queue:

• Running an evaluation

• Running a trace evaluation

• Running GEPA optimization

• Ingesting a large trace batch

For async operations:

• The API returns a handle (an evaluation ID, optimization ID, etc.)

• The work runs in the background

• You poll for status, subscribe via SSE, or call wait_for_completion(...) in the SDK

SDK polling helpers

The SDK has helpers like:

result = client.evaluations.wait_for_completion(eval_id, timeout=300)

Async-equivalent:

result = await client.evaluations.wait_for_completion(eval_id, timeout=300)

Queue depth and ECU

Async work queues per-tenant. Most queues drain in seconds; large GEPA optimizations or large evaluation batches may take minutes.

Failure semantics

Async runs that fail report a failure status with an error code. The SDK exception classes (APIError, NotFoundError, etc.) cover both immediate and async errors.

Where to next

• Python SDK reference

• Evaluations

• Trace evaluations