Agents
Start AI-driven development work directly from a natural language prompt. Choose between a spec-first or code-first approach and let agents work inside a TDX-secured sandbox.
The Agents page is the starting point for AI-driven work inside a project. Rather than navigating to a task and opening a worktree manually, you describe what you want to build in plain language, choose how you want to approach it, and the agent takes it from there, either by thinking through a spec with you first, or by going straight into code inside a secure sandbox.
It is accessible from the project sidebar under Agents.
The Start Screen
The main area presents a single prompt input:
Describe the app, feature, or change you want to make…
Below the input are two mode buttons that control how the agent approaches the work:
Spec/PRD Mode
Selecting SPEC/PRD opens a feature ideation session. The agent reads your project's code index for context and works with you in a structured chat to define the spec, scope, approach, edge cases, and implementation plan, before any code is written.
This mode is useful when:
- You have a rough idea but haven't thought through the details yet
- You want to align on approach before committing to an implementation
- The change is large enough that getting the spec wrong would be costly to unwind
At the end of a Spec/PRD session you can Save the spec as a document, Save as New to create a separate copy, or Generate Tasks to turn the spec directly into actionable tasks on the project board.
Code Mode
Selecting CODE goes straight into an active coding session inside a new worktree. A branch selector appears so you can choose which branch to base the work on. The agent provisions a TDX-encrypted sandbox, clones the repository, and begins working on your prompt immediately.
This mode is useful when:
- The change is well-understood and you want implementation without a planning step
- You are making a targeted fix, refactor, or addition
- You want to iterate quickly and review the diff yourself
Every Code Mode session runs in a TDX Sandbox, the same confidential compute infrastructure used by all worktrees in Origin. See Sandboxes for details on what that means for isolation and security.
Once you have filled in the prompt and selected a mode, click START.
Spec/PRD Session
Inside a Spec/PRD session the layout splits into two panels:
Left: conversation: The agent greets you and opens a feature ideation chat. Describe the feature, ask questions, push back on suggestions, the agent maintains context of the full conversation and refines the spec as you go.
Right: context: The project's code index is loaded as a reference document. This is the same indexed representation of your repository that agents use during code execution. The agent draws on it to ground suggestions in your actual codebase rather than producing generic output.
The bottom bar shows the active model (Claude Sonnet 4.6) and the current mode (Feature Ideation).
When the spec is ready, use the actions at the bottom right:
- Save: save the spec document to the current project
- Save as New: save as a new document without overwriting the current one
- Generate Tasks: convert the spec into tasks on the project board (see below)
Generating Tasks from a Spec
Clicking Generate Tasks (or the inline generate tasks for this doc button in the chat) shifts the agent into Tasks Planning mode. Before creating anything, the agent may ask one or more clarifying questions to make the tasks accurate, for example, asking where the feature will live or who the primary users are. You can answer with the provided options or type your own.
Once it has enough context, the agent generates a structured task breakdown directly in the chat:
- A parent task representing the overall feature or change
- Numbered subtasks, each with a defined Scope and a set of Acceptance Criteria
The tasks reflect the specific requirements from the spec, a more detailed or constrained spec produces more targeted tasks, while a broad prompt produces broader ones. The number of tasks varies based on complexity.
After generation, the chat confirms how many tasks were created (e.g. "8 tasks created successfully") and the full list appears in the right panel with all scope and acceptance criteria visible. The tasks are simultaneously added to the project board and are immediately available under Tasks for assignment, prioritisation, and execution.
Code Sessions and Recent Worktrees
The start screen shows two persistent sections below the prompt:
Resume Existing
Lists recent feature chats (Spec/PRD sessions) you can return to. Selecting one reopens the conversation and its saved context so you can continue where you left off or refine the spec further.
Recent Worktrees
Lists active and recently run worktrees started from the Agents page. Each entry shows:
- The session name (e.g. Code Mode)
- A TDX Sandbox badge confirming the worktree ran inside a confidential compute environment
- A timestamp showing when it was last active
Clicking a worktree entry reopens it. If it is still running, you return to the live session. If it has been archived, you can review its output and changes.
Left Panel
The left panel provides quick access to common actions without going through the prompt input:
- START: returns to the main start screen
- NEW SPEC CHAT: opens a blank Spec/PRD session immediately
- NEW CODE SESSION: opens a blank Code Mode session immediately, skipping the start screen
Previously saved spec documents are listed in the panel under FEATURE / PRD DOCS as they accumulate, so you can reopen any past spec without searching.
Task Mind Map
View every task in the current project as an interactive mind map. Navigate the project's task tree, spot priorities, and open tasks directly from the map view.
Project Features & Context
Add and manage supporting documents in the Context section to help AI agents understand your project's needs. Upload specs, requirements, and reference files to improve agent accuracy and output quality.