Pre-release guide
Design a recoverable desktop run.
Plan a computer-use workflow around application readiness, bounded control, durable artifacts, and deliberate recovery.
Planning a pilot? Share the workloadAdept AI is pre-release. Final methods, limits, supported images, retention, and commercial terms are confirmed only in a written pilot scope.
Define the workload
Describe the application version, input artifacts, desired output, expected run length, and the evidence an operator needs to review.
Confirm capabilities
A pilot environment should report the controls, limits, application image, and recovery features actually available before an agent begins work.
Create the environment
Provision a clean application-ready desktop with a bounded workspace and an idempotent request identifier.
from adept import Client
client = Client()
environment = client.environments.create(
"kicad",
name="Board review",
idempotency_key="board-review-v1",
)
environment.wait_until_ready()
capabilities = environment.capabilities()
print(capabilities)Connect the control loop
Observe the screen, send pointer and keyboard input, and move approved files through explicit operations.
Checkpoint useful progress
Save inside the application, create a checkpoint, and verify that persistence succeeded before a risky step.
- Save the active document in the application.
- Create the checkpoint request.
- Wait for a durable success response.
- Record the checkpoint identifier with the agent trace.
Stop, inspect, or recover
Collect the intended artifact, preserve the run record required by the pilot, and either stop cleanly or restore a verified checkpoint.
Proposed control groups