The AI Agent
ZeroSphere is a proactive development environment. Its agent writes code, compiles it, launches your application, interacts with the running UI, and fixes what breaks — automatically, without you needing to describe the problem.
What the Agent Can Do
Write and edit code Creates files, modifies existing code, and refactors across multiple files. It reads your full codebase, not just the file currently open.
Run commands Executes builds, test suites, and shell commands. Reads the output and acts on it.
Launch and operate your app Spins up your application in an isolated virtual display. Clicks buttons, scroll app, and navigates UI flows. The agent sees raw pixels — no DOM, no accessibility tree.
Fix what breaks When a test fails or a UI flow breaks, the agent traces the failure back to the responsible code and patches it.
How a Task Runs
You give the agent a task:
"Add authentication to the login route and test that it works"
The agent:
- Reads your codebase and understands the structure
- Writes the implementation
- Compiles the project
- Creates a virtual display session
- Launches your app inside it
- Clicks through the login flow
- Catches failures and fixes them
- Reports back when done
Permissions
Every action outside normal code editing goes through the permission system.
| Mode | Behavior |
|---|---|
ask_each_time | Agent pauses and waits for your approval |
run_always | Agent acts immediately without prompting |
let_model_decide | Agent evaluates risk and asks only when it judges necessary |
You can change permission settings mid-task. The agent will not act outside your permitted scope without asking.
See Permissions for the full breakdown.
The Virtual Display
When the agent launches your app, it opens in an isolated virtual display — a leased VDD monitor that Windows treats as a real connected screen, separate from your desktop. The live feed is visible in the agent panel.
- The agent cannot see or interact with your main desktop
- You can watch every click and keystroke in real time
- Sessions expire after 20 minutes of inactivity and clean up automatically
- You can take control at any point
For apps that cannot run in a virtual display, the agent can take control of your real primary display instead. A floating STOP button is always visible so you can interrupt.
See Virtual Display for details.
Context the Agent Uses
The agent works from:
- Conversation history in the current session
- Files you attach or reference explicitly
- Command output and test results from the current task
- Live screenshots from the virtual display
Context does not persist between sessions. Each new session starts fresh.
Planning Mode
Before writing any code, the agent creates a full implementation plan and presents it to you. Nothing executes until you approve the plan. After approval, it executes step by step and saves a completion report when done.
Planning artifacts are saved to %USERPROFILE%\.zerosphere\plan/.
Switch modes via the mode selector in the agent panel.