Review is a distinct state
A finished worker enters a review step. Acceptance requires evidence, keeping worker completion separate from supervisor acceptance.
SUPERVISED AI WORKFLOWS
Second Cursor
Automation with a review step.
The problem
An AI worker saying “done” does not establish that the task is finished. Useful automation needs visible progress, reviewable evidence, limits on retries, and a way to stop.
Built for developers coordinating AI coding tasks and experimenting with scoped Linux desktop actions.
What I built
I built an MCP-based supervisor that tracks tasks, collects review evidence, and supports bounded correction loops. Separate desktop-control infrastructure checks the target and pauses when human intervention invalidates the working context.
Define the task
Scope + completion criteria
Run & track
Worker output + visible task state
REVIEW GATEDoes the evidence support completion?
Inspect the resultHuman review
ACCEPTComplete
CORRECTBounded retry ↰
Cancellation remains a separate control.
THE WORKFLOW
From the first step to the result.
- 01
Define
Set a task and completion criteria before dispatching a worker.
- 02
Track
Follow task state and inspect the resulting files or other evidence.
- 03
Review
Accept a result only after review, or request a correction within the task’s budget.
- 04
Resolve
Finish an accepted task, report an unresolved issue, or cancel the work.
ENGINEERING DECISIONS
The details
behind the work.
Bound the correction loop
Correction limits vary by task type. A task cannot quietly retry forever: exhausted budgets become an issue to resolve.
Separate supervision from desktop input
Scoped Linux controls check window/process identity and observation freshness. Human input can pause automation; resuming requires a fresh observation and explicit action.
CURRENT SCOPE
Where it stands.
Task tracking, evidence review, bounded corrections, cancellation, and scoped Linux desktop-control mechanisms are implemented. Current end-to-end demonstration is pending. Desktop controls are tied to a constrained Linux/Hyprland setup.
Supporting verification notes
The illustrated flow represents source-level mechanisms, not a recorded successful AI run. Desktop controls have monitor and compositor constraints. Cooperative worker instructions are distinct from enforced process boundaries; complete security isolation or universal autonomous desktop operation is not claimed.
APPLYING THE APPROACH
Discuss practical automation.
These patterns are useful for document processing, task supervision, and AI-assisted workflows where people need to inspect the result before it becomes final.
Discuss a similar project Opens your email application.