Copilot Cockpit

01 Problem

Powerful AI agents still need memory, boundaries, and review points.

Prompt-by-prompt development makes it difficult to keep priorities, research, execution, and approvals connected. Autonomous loops can also hide too much: what is running, why it was chosen, what it changed, and where a person should intervene.

02 System

A control plane inside the editor where the work already happens.

Copilot Cockpit adds durable planning and execution surfaces to VS Code. Todo Cockpit holds discussion and approval state; Tasks represent focused runs; Jobs coordinate ordered steps and checkpoints; Research tracks measured improvement. Optional specialist agents can handle bounded slices without replacing human closeout.

• 01
  Approval firstPlanning artifacts stay separate from executable tasks until the user deliberately promotes them.
• 02
  Different tools for different workTasks, Jobs, and Research remain distinct because one prompt, an ordered workflow, and a benchmark loop have different control needs.
• 03
  Repo-local by designConfiguration and working state remain attached to the workspace, keeping the operating context close to the code.

03 Role

Product definition shaped through intensive real-world use.

I developed the product model, interaction system, extension architecture, workspace persistence, onboarding, and agent-control concepts while using the tool against active software work. The result reflects a repeated pattern in my projects: make a complex technical process easier to inspect and operate without hiding its important decisions.

04 Evidence

A working extension with a deliberately layered workflow.

The public repository includes Todo Cockpit, Tasks, Jobs, Research, workspace settings, onboarding, scheduling, and optional integrations. The project documentation clearly separates stable workflow primitives from experimental features instead of presenting every idea as equally mature.