Complete your first agent task
Complete a small agent task, inspect its result, and continue with a focused follow-up.
Last updated 16 September 2026
On this page
Your first task should be small enough to review confidently. This walkthrough takes you from a clear request to a result you have checked, so you learn both how to ask an agent for help and how to decide whether its work is ready.
Before you start
Add a local project and connect your coding provider. Choose a familiar project and one bounded outcome: explain a module, correct one label, or fix one reproducible issue. Avoid combining unrelated changes in your first request.
1. Set up the conversation
Select the project in the sidebar and open New Agent (⌘N by default). Check the project, model, reasoning effort, and access setting in the composer. Available choices depend on your provider. Use the conversation mode control to choose planning when you want to agree on an approach before implementation.
Demo example: the new-agent composer for Northstar Client Portal. The model, effort, access, and conversation controls sit below the message area; the project and Solo controls sit beneath it. A running demo preview is open on the right. This is an unsent task, not evidence of completed agent work.
2. Describe a result you can check
Include the desired change, where it belongs, what should stay unchanged, and how to check success. Mention relevant files or attach a screenshot when that makes the request clearer. For example:
In the client overview, change the heading from ‘Welcome’ to ‘Your projects’. Keep the existing layout and styling. Tell me which file changed and how to check the result.
If you prefer to start without changes, ask the agent to explain the main entry point and run command, and explicitly say not to edit files. That still teaches you how project context and follow-up questions work.
3. Follow progress and answer decisions
Send the request and read the activity in the conversation. When the agent asks a question, answer the specific missing detail. If it presents a plan, check that the proposed scope matches your request before authorizing implementation. Read permission prompts before approving an operation.
Keep using the same conversation for clarifications about this task. Sending a duplicate task to another agent while the first is running can produce overlapping work.
4. Inspect the result
Read the final explanation, including any limitations or checks the agent could not perform.
For an implementation task, open Changed Files and inspect the diff. Check that it contains the intended change and no unrelated edits.
Run the relevant behavior in your project or Preview. For the heading example, open the client overview and check the exact text and layout.
Follow the relevant What to check items when a checklist is available. Ask for a focused fix if a check fails.
An agent finishing its turn is a useful handoff point. Your review establishes whether the actual result meets the request.
Continue or deliver
If a detail is wrong, describe the observed result and the expected result in a follow-up, such as ‘The new heading appears on desktop, but mobile still shows Welcome.’ Once the change is ready, use your Git or Ship workflow to deliver it. Finishing a conversation does not itself mean the changes have been committed, pushed, or merged.
If the task cannot start
Check the selected project and provider sign-in status. If the agent is waiting for a question, plan decision, or approval, respond to that item instead of resending the original request. If it reports an error, keep the message and any relevant output so you can address the cause.
