Autonomous mode

Autonomous mode lets the agent own the outcome of a task from start to finish. Instead of iterating with you step by step, the agent builds a plan, delegates work to specialized sub-agents, and opens a pull request or merge request automatically when the work is complete.

Autonomous mode is distinct from declarative workflows. Autonomous mode lets the agent decide its internal process for one task. A workflow gives the runtime a reusable graph with explicit steps, agents, handoffs, loops, waits, and completion conditions; it can run in the background while you continue the parent conversation.

Toggle Autonomous in the chat input bar, either before your first message or at any later point in the session, to hand off without starting over. Autonomous mode is off by default — when it's off, you work with the agent collaboratively in the default mode.

In autonomous mode, the agent selects the model automatically — you cannot choose the model yourself.

When you submit a task with autonomous mode enabled, the agent follows a structured process:

1. Planning — The agent breaks down the work into steps, generates requirements, and defines acceptance criteria.
2. Execution — Specialized sub-agents handle each step of the plan — analyzing the codebase, writing code, and verifying changes before moving forward.
3. Completion — The agent opens one or more pull requests or merge requests with detailed descriptions of the changes, implementation approach, and any trade-offs considered.

Throughout this process, the agent operates in an isolated sandbox where it clones the repository and works on the task.

Autonomous mode works well for tasks where you can clearly describe the desired outcome and the agent can execute independently:

• Implementing a well-defined feature across one or more repositories
• Fixing bugs with clear reproduction steps
• Writing tests for existing code
• Refactoring code to follow established patterns
• Updating documentation or configuration files

For exploratory work, design discussions, or tasks where you want to iterate closely with the agent, use the default mode instead.

You can monitor autonomous tasks from the session view. The agent provides updates as it moves through each phase. If the agent needs clarification during execution, the task moves to a Needs attention state and waits for your input.

The agent won't always open a pull request — for some tasks it may provide answers, analysis, or suggestions directly in the chat. When it does open a PR:

1. Check the code changes and implementation approach
2. Leave comments with feedback or suggestions
3. The agent automatically addresses your comments and pushes updates

You can also use PR commands to direct the agent:

• /kiro all — Addresses all comments from all reviewers across the entire PR
• /kiro fix — Addresses all comments within a specific conversation thread

Use PR feedback to teach the agent your team's patterns. Comments like "always use our standard error handling" or "follow our naming conventions" help the agent learn and apply those patterns to future work. See steering for more ways to guide the agent.

Learn more about working with the agent on GitHub.