Agent
Plans, todos, and subagents
How the agent plans larger jobs, tracks its steps, asks you questions, splits work across subagents, and shows data in Slates.
For work bigger than a quick edit, the agent organizes itself. It can write a plan for you to approve, keep a todo list as it goes, stop to ask you a question, hand parts of the job to subagents, and build a Slate when the answer is data to look at.
Plans
Switch to Plan mode and describe the job. The agent researches the codebase and writes a plan instead of editing code. The plan appears in chat as a card marked Plan ready.
- Click View Plan to open it. You can edit the plan before you approve it.
- Click Implement to switch the chat to Develop and start the work.
- The card shows progress as steps finish, for example In progress · 3/7 done. If a run stops early, click Continue.
Plans are saved as markdown files in .velork/plans/ in your workspace. A chat has one active plan at a time. When you change direction, the agent edits that plan rather than writing a new one.
Todos
On jobs with several steps, the agent keeps a todo list in the chat under Tasks. Each item shows whether it is pending, in progress, or completed, and the header counts what's done (3/5). Items the agent cancels drop off the list. The list updates as the agent works, so you can see where it is at a glance.
This list belongs to the chat. It is separate from the task board.
Questions
When the agent hits a choice it can't make from your project or your message, such as a missing API key or a decision with different outcomes, it asks you in the chat with options to pick from. Its suggested option is marked Recommended, and some questions let you choose any that apply or type your own answer. With several questions, use Back and Next, then Submit. Dismiss (Escape) skips them.
With Autonomous on, a supervisor answers for you.
Subagents
The agent can start subagents for independent parts of a job, such as researching two areas of a codebase at once or building separate pieces in parallel. Each subagent runs in its own session with a fresh context and reports back to the main agent when it's done.
Subagent runs appear as cards in the chat. Click one to open its full session, with every tool call and edit, then click Back to return.
There are three kinds:
| Kind | What it can do |
|---|---|
| General | Edit files and run commands |
| Explore | Read and search the codebase and canvas, without editing |
| Review | Review the branch's changes and report must-fix issues and nits, without editing |
You don't need to ask for subagents. The agent uses them when splitting the work helps, and you can ask it to run parts in parallel.
Slates
When the answer is something to look at rather than read, such as metrics, comparisons, rankings, audits, or benchmark results, the agent builds a Slate. A Slate is a live page with tables and charts. It opens as a tab in the agents panel, or as an editor tab in the IDE layout. Click Open Slate on its card to bring it back.
Compare bundle sizes for every route before and after this branch.