Working with the agent

Kiro Web's (Preview) default mode is collaborative — you work with the agent interactively, iterating on a problem together. The agent helps you think through approaches, write code, and open pull requests when you're ready.

For tasks where you want the agent to own the outcome end-to-end, see autonomous mode.

1. Go to app.kiro.dev and click New session
2. Select one or more repositories using Select repo at the bottom of the chat input
3. Describe what you want to work on

If you don't select repositories upfront, you'll be prompted to choose them when you ask the agent to work on something.

In the default mode, you drive the conversation. Chat with the agent to:

• Discuss implementation approaches and trade-offs
• Ask questions about your codebase
• Get the agent to write or modify code
• Iterate on the implementation with feedback
• Ask the agent to open a pull request when you're satisfied

The agent has access to web search, context from your repositories, and learnings from previous code reviews to provide informed responses.

When you're happy with the changes, ask the agent to open a pull request. The agent creates a feature branch, commits the changes, and opens a PR with a description of what was done.

You can continue to provide feedback on the PR. The agent responds to your comments and pushes updates.

You can work across multiple repositories in a single session. When you select repos, the agent clones them into its sandbox and can coordinate changes across all of them.

Each session has a single conversation thread. You can:

• Steer the agent on implementation approach at any time
• Provide additional requirements or constraints
• Ask the agent to do more work after reviewing results

To work on a different topic, start a new session.

Sessions expire after 90 days, at which point chat messages and logs are deleted. Pull requests, code changes, and conversations on GitHub are not affected.

Kiro Web can search the web to access current information in real-time, enabling up-to-date answers about topics that may have changed since the model's training data was created.

Web search is designed to not reproduce meaningful chunks of text and cannot access pages behind paywalls or authentication. Web searches are logged in the session with citations, so you can see what sources the agent referenced.