Autopilot Positioning
Click a card to see the characteristics of each mode / permission level. The "Autopilot" concept exists in both VS Code and the CLI.
VS Code
๐Ÿ›ก๏ธ
Default Approvals
โšก
Bypass Approvals
๐Ÿค–
Autopilot
Preview
Same autonomous-
execution concept
CLI
โœ๏ธ
Suggest
๐Ÿ’ฌ
Interactive
default
๐Ÿ“‹
Plan
๐Ÿค–
Autopilot
NEW
VS Code: Autopilot (Preview) โ€” Fully automatic approvals + auto-answers to AI questions. The highest-level mode that runs autonomously until the task is complete.
Standard Mode vs Autopilot Flow
In Autopilot mode, the AI repeats an autonomous loop and minimizes interruptions to the user.
๐Ÿ’ฌ Standard Mode (with approvals)
๐Ÿ‘ค
Send instruction
โ†“
๐Ÿค–
Generate suggestion
โ†“
๐Ÿ‘ค
Approve / request changes
โ†“
๐Ÿค–
Output revised proposal
โ†“
๐Ÿ‘ค
Approve again
โ†“
๐Ÿค–
Confirm completion
๐Ÿค– Autopilot (Autonomous)
๐Ÿ‘ค
Initial instruction only
โ†“
๐Ÿ”„ Autonomous loop (internal steps are not billed)
๐Ÿ“ Edit files
โ†“
๐Ÿงช Run tests
โ†“
๐Ÿ”ง Fix errors
โ†“
๐Ÿ” Re-run tests
โ†“
โœ… Completion check
Autopilot Stop Conditions
Autopilot stops automatically when any of the following conditions are met. Even in fully autonomous execution, humans always remain in control.
--max-autopilot-continues simulation
1 50 10
Auto-stop after 10 steps. A good setting for moderate-complexity tasks.
--allow-all flag
ON
Grants access to all tools, paths, and URLs. Recommended for Autopilot.
Request Count Comparison
Interactive mode generates a request for each human action. Autopilot's internal loop steps are not billed (shown in grey).
๐Ÿ’ฌ Interactive (5 exchanges)
๐Ÿ‘ค Human sends +1
๐Ÿ‘ค Human sends +1
๐Ÿ‘ค Human sends +1
๐Ÿ‘ค Human sends +1
๐Ÿ‘ค Human sends +1
Total: 5 requests
๐Ÿค– Autopilot
๐Ÿ‘ค Human sends +1
Total: 1 request ← before execution
Internal tool calls (file reads/writes, shell execution, etc.) occur within Autopilot's autonomous loop but are treated as non-billed steps. Only AI turn transitions (= API requests) are charged.