Interactive guide — 01 # The Autonomy Dial Four observable things, not vibes: who initiates the work (you, a ticket, a queue, or the job itself), where the approval gate sits (per suggestion, per action, per result, or on policy), the blast radius if it goes wrong, and how you verify the work (reading suggestions, reviewing diffs, auditing logs, or managing outcomes). Pick a level — or use ← / → when the dial is focused. Exemplar chips link to live spec sheets. You delegate a ticket; it returns a finished artifact — usually a PR. **Initiated by**: A ticket, mention, or assignment — often outside your editor **Approval gate**: Task acceptance, then the review of the result **Blast radius**: A branch, a sandbox, a bounded scope of work **You verify by**: Reviewing the PR, run logs, and test output In the catalog at this level [DevinCloud & autonomous agents](https://www.agentlist.io/systems/devin)[Factory DroidsCloud & autonomous agents](https://www.agentlist.io/systems/droid)[JulesCloud & autonomous agents](https://www.agentlist.io/systems/jules)[OpenHandsCLI & terminal agents](https://www.agentlist.io/systems/openhands) Next level unlocks when — Let it run continuously — a queue of tasks, not one assignment at a time. FAQ ## Frequently asked ### What do the autonomy levels actually measure? Four observable things, not vibes: who initiates the work (you, a ticket, a queue, or the job itself), where the approval gate sits (per suggestion, per action, per result, or on policy), the blast radius if it goes wrong, and how you verify the work (reading suggestions, reviewing diffs, auditing logs, or managing outcomes). ### Where does a 'coding agent' sit vs a 'general agent'? Category and autonomy are different axes. A CLI coding agent asking before each shell command is L1; a cloud agent that takes a ticket and returns a PR is L2; a customer-facing digital employee on shift is L4. 'General' vs 'coding' describes the domain — the dial describes how much you delegate. ### Can one product span multiple levels? Yes — most agent products do. GitHub Copilot started at L0 autocomplete and now runs an L2 coding agent that opens PRs. The dial describes modes of operation, not companies. Product pages in the catalog show which interfaces and capabilities a system ships today. ### Does higher autonomy mean better? No — it means more is delegated. Each step up trades per-step control for leverage, and raises the stakes on verification: at L0 you proofread, at L2 you review PRs, at L4 you manage outcomes. The right level is the highest one your verification story can actually support. --- Source: [The Autonomy Dial — five levels of AI agent delegation | agentlist.io](https://www.agentlist.io/learn/autonomy). This is the public page rendered as Markdown; interactive controls require the website.