Safe control loops for desktop agents
How to structure observations, bounded actions, file movement, and human approvals around real software.
Keep control primitives small
Screen observation, pointer input, keyboard input, file transfer, and application-scoped execution should remain separate capabilities. This makes side effects visible and lets the client apply different retry and approval rules to each operation.
Batch only actions whose combined result can be inspected safely. Long, opaque desktop scripts make it harder to locate partial progress when a timeout or dialog interrupts the sequence.
Treat limits as part of the protocol
File-size bounds, cell deadlines, screenshot dimensions, concurrency, and one-operation-at-a-time rules are not implementation trivia. Agents should discover or configure them before work and adapt the task plan accordingly.
The researched Isle interface currently documents 4 MiB individual transfers, 120-second execution cells, bounded output, and exclusive desktop operations in new supported images (source).
Humans retain authority
An agent operating CAD or EDA software can change valuable technical artifacts, but it should not be treated as an authorized engineering approver. Fabrication, release, procurement, safety, and access decisions remain with qualified people.
Adept AI’s proposed controls support checkpoints and traceability; they do not prove that an action, result, or model conclusion is correct. Verification must match the consequence of the work.