Skip to main content
AI Work is the agent pane you can use while navigating ReasonOS. It can read and act on your project’s work, using your permissions. You do not have to open a workspace first, but the agent still runs on a branch server.

​
Open the pane

Use the robot icon + Work button in the top bar, or beside Open workspace on a project’s overview. Its accessible name is Work with AI. The pane is available to signed-in people in organizations with the core ReasonOS product; it is not a separate item in Tools. The editor’s Tools → AI action opens the workspace’s own AI panel, not this app-wide pane. The project you are viewing is already selected. From a workspace, the current branch comes with it. From a project page without a branch context, the agent uses the project’s default branch. If no project is in context, choose one in the pane. Check the Project and Branch shown before sending work. You can also open the pane with a record in context:
  • Plan with AI on a cycle prepares a planning prompt for that cycle.
  • Review with AI on a proposal prepares a review prompt for that proposal.
  • On an issue, Run with agent is a different, direct action: it saves an instruction and asks the issue’s branch agent to start. See Working on project issues.
Cycle and proposal prompts are editable drafts, not automatic model calls.

​
Start AI work

  1. Check the project and branch.
  2. If no server is running, choose Start a server and open to authorize starting one. This uses a branch-server slot. Choose Not now to leave it stopped.
  3. Review Agent and AI model. Use the displayed defaults, or choose an override for a new conversation. ReasonOS (built-in) is the built-in agent; Claude Code, Codex and Gemini CLI are other agent software choices when available on that server.
  4. Choose Open the conversation. A previously opened AI Work conversation for you on this project’s branch is resumed with its own agent and model; a setup override does not replace those of an existing conversation.
  5. Review or write your message, then send it.
Opening the pane, opening the conversation and changing setup selections do not send a model prompt. Starting a server may consume infrastructure resources even before you send a message. For the built-in agent, Account or key identifies the selected model’s access source: your AI account, a project API key or an organization API key. It is not a price or a subscription tier. CLI agents use their vendor account configuration rather than the built-in agent’s project-key selection. See AI accounts and keys for the difference between the agent, the model and the credentials that pay for the call.

​
Defaults and overrides

The setup explains whether it is using branch settings, your last choice on this browser, or a new override. Model choices use readable names and are grouped by provider. A CLI can use its own default when no model list is available for its account—for example, Codex default for a ChatGPT sign-in. Use branch defaults forgets the last model choice stored in this browser for new conversations. It does not delete keys, change the branch’s settings, or change existing conversations. This preference is not an account setting that follows you to every browser. If a remembered agent is unavailable or its model is incompatible, choose an available agent/model or use branch defaults. The pane does not silently pretend to use a different choice.

​
Keep your context while navigating

The pane stays with you across pages. Hiding it, or pressing Escape to hide it, does not stop the conversation. An open conversation remains on its original branch; visiting another page does not move that branch server or check out a different branch over someone else’s work. When a record action or project selection asks to change context, the pane offers Use this context or Keep current conversation. While the agent is working, wait for it to finish before switching. Unsent drafts are kept in the current tab; do not rely on them surviving a reload or closing the tab. Clear record context removes the selected issue/cycle/proposal context, not the conversation’s saved messages.

​
What the agent can do

The agent uses ReasonOS’s project-scoped tools to read issues, update statuses, comment, and work with other project records your account can access. For example: “Read issue 42, summarize its subtasks, and comment with a plan.” Actions still require the same permissions and product access as the UI. An agent can also request work on another branch already linked to an issue, using a comment you wrote as the instruction. A stopped target needs explicit permission to start a server. An accepted request means the branch received it, not that the work is finished. Read the issue’s activity for the outcome. See Work tools over MCP.
AI Work belongs to the managed ReasonOS application. The standalone local editor does not provide this app-wide pane. Available agents and models depend on the deployment, installed agent software, linked accounts and configured keys. These guides describe the current application; an older deployed build may not yet show the same labels or actions.