Grind Mode

Tackle large QA tasks with parallel agents that plan and execute over hours.

Grind mode is designed for large QA tasks that require planning, coordination, and extended execution time. Instead of a single agent processing your prompt, Grind uses a planner that breaks your task into subtasks and delegates them to parallel workers.

When to Use Grind

How Grind Works

Phase 1: Planning

The agent investigates your codebase — reading files, running builds, checking tests — and produces a structured plan with phases and subtasks.

The plan is presented for your review before any code changes are made.

Plan Approval

You control when and how plans are approved:

• Manual approval — Review the plan and approve or reject it
• Auto-accept — Plans execute automatically without waiting for approval

When a plan is awaiting approval, you can:

• Approve — The agent begins execution
• Reject — The agent revises its approach
• Send a message with additional context or corrections

Phase 2: Execution

Once approved, the planner delegates subtasks to parallel workers:

• 2-8 workers run simultaneously on different parts of the codebase
• Each worker is scoped to specific directories to avoid conflicts
• Workers can be assigned different model tiers based on task complexity
• Sub-planners can recursively delegate for deeply complex tasks (up to 3 levels)

Verification

After all workers complete, the agent:

1. Runs the build to verify compilation
2. Runs tests and compares against the baseline
3. Runs linters and checks for regressions
4. Commits and pushes changes

Enabling Grind Mode

1. Open the Agent page in the dashboard
2. Toggle Grind in the input bar
3. Select a time limit (e.g., 3 hours, 5 hours, 10 hours, or until done)
4. Choose your plan approval preference (manual or auto-accept)
5. Send your prompt

Sandbox Environment

Grind sessions run in an isolated sandbox with a full desktop environment. The Sandbox Panel gives you:

• Desktop view — Live stream of the agent's virtual desktop via VNC
• File changes — Real-time view of modified files
• Plan phases — Track progress through the execution plan
• Terminal — See command output

This is especially useful for E2E testing and UI validation, as the agent can interact with web applications visually and verify behavior in a real browser.

Tips

• Be specific — Detailed prompts produce better plans. Include what to test, expected behavior, edge cases, and constraints.
• Start with auto-accept off — Review the first few plans manually to calibrate quality, then switch to auto-accept for trusted workflows.
• Use for coverage goals — Grind is ideal for "add tests for all untested modules" or "fix all flaky tests in this repo" type tasks.