Autonomy SDK

Autonomy turns governed experience into a deployable ability. The primary input is an application manifest: a human-readable build and delivery intent with no service URLs, credentials, or runtime topology.

Load an application manifest

from pathlib import Path

import yaml
from sx_autonomy import application_manifest_from_dict

manifest = application_manifest_from_dict(
    yaml.safe_load(Path("application.yaml").read_text())
)

The parser rejects ambiguous build references, invalid workload connections, secret values, and incomplete applications before any remote work begins.

Train the build graph

async with sx.connect(project_id) as client:
    operation = await client.train(manifest)
    print(operation.ref)
    outcome = await operation

client.train(...) executes only the manifest build graph. It returns a durable Autonomy program operation whose result carries governed outputs and evidence.

Run the full promotion program

async with sx.connect(project_id) as client:
    operation = await client.deploy(manifest)
    print(operation.ref)
    deployment = await operation

client.deploy(...) is the preferred golden path. It resolves or builds the application, evaluates the exact graph, qualifies it for the target, mints task-specific approvals, and rolls out the declared deployment group. A failure at any gate stays explicit.

Run a bounded improvement loop

from datetime import timedelta

task = await client.experience.task(task_ref)
operation = await client.autonomy.improve(
    task,
    max_rounds=4,
    timeout=timedelta(hours=2),
)
result = await operation

The round count and timeout are authority boundaries, not suggestions. The campaign can form hypotheses, gather evidence, and iterate inside them; widening spend, changing intent, or promoting still requires accountable authority.

Approve exact evidence

For a custom promotion flow, use the values produced by Worlds and qualification directly:

approval = await client.autonomy.approve(
    evaluation,
    qualification,
    supervision_requirements,
)

The approval binds the exact evaluation, whole-application qualification, task, and supervision requirements. It is not a general permission to run a model.

Use the right surface