The golden path

The golden path is the shortest complete robotics story: define one task, gather governed experience, evaluate and improve an exact application, deploy it under supervision, and return executed outcomes to the catalog.

Designed path, measured status

The verbs in this path are implemented at different surface altitudes, but the customer journey has not yet been traversed end to end without intervention. Check the generated surface status and the canonical journey gaps before running or promising the whole loop.

The lifecycle

Define the task in Experience

Publish one immutable task contract. Its exact reference follows every capture, episode, evaluation, application approval, deployment, and returned outcome.

Gather governed experience

Capture against the validated task, preserve raw input, register canonical RRD episodes, and freeze the exact selection used downstream.

Iterate in Worlds

Use sessions or generations for learning and exploration. Use a held-out evaluation of the exact application for promotion evidence.

Build the ability in Autonomy

Train from an immutable corpus snapshot, inspect the run evidence, and assemble the exact RobotApplication graph. A bare checkpoint is not the deployable unit.

Promote through Fleet

Qualify the application, bind task-specific approvals, deploy to a canary group, and supervise it. An unanswered escalation ends in an explicit safe-stop.

Close the loop

Return only executed actions, corrections, and outcomes to Experience with their deployment, approval, and task-lease lineage. The next corpus can now improve on real evidence.

Let an agent drive the routine work

After sx project connect, give your coding agent an accountable goal and a budgeted boundary:

Move task <task-ref> toward a supervised canary using the Sentient-X MCP tools.
First inspect the current resources and cite every exact reference. Use the canonical
Experience → Worlds → Autonomy → Fleet path. Do not widen the task, spend budget,
start physical motion, or promote beyond canary without asking. Reattach to durable
operations and report success only from terminal receipts and governed evidence.

The agent should return a trail you can audit:

exact task, corpus, application, qualification, approval, and deployment references;
durable operation references and terminal states;
evaluation or RRD evidence;
every refusal and the next safe recovery action.

The same model in Python

The SDK keeps pillar ownership visible and returns typed durable operations.

from uuid import UUID

import sx
from sx_worlds import WorldEvaluation


async def evaluate(project_id: UUID) -> WorldEvaluation:
    async with sx.connect(project_id) as client:
        policy = await client.autonomy.policy("policy:pick-mug-v12")
        operation = await client.worlds.evaluate(policy)
        print(operation.ref)  # save this; disconnecting does not cancel work
        return await operation

For the complete typed capture-to-deploy contract, read the executable golden-path test.

One promotion rule

Promote evidence, not confidence. A result moves forward only when its owning gate can prove the exact identity that passed. Keep exploration flexible inside a stage; keep handoffs strict.