Agents
Origin Agent, Agent Sessions, Studio agent chats, and subagent model routing, the related but distinct uses of "agent" in ORGN.
CDE and ORGN Studio use the word "agent" in several places. This page separates them so you know which surface (or which setting) to reach for.
Three surfaces at a glance
| Origin Agent | Agent Sessions (CDE) | Agent chats (Studio) | |
|---|---|---|---|
| What | In-IDE assistant — Q&A, edits, and tool use on demand | Desktop UI for browsing long-running agent sessions | Browser sidebar for autonomous build agents on tasks |
| Product | CDE desktop | CDE desktop | ORGN Studio browser |
| Where | Chat panel | Agent Sessions Welcome page; chat auxiliary bar | Project sidebar — labeled Agent Chats |
| Open via | Chat panel in the workbench | Command palette: Open Agent Sessions Welcome | Appears when build agents are running on tasks |
| Best for | Day-to-day work in your current workspace | Review, switch, and project focused session views | Monitor autonomous agents launched from the task board |
These are not interchangeable. Origin Agent is the in-IDE assistant in your current workspace. Agent Sessions is a separate session browser. Agent chats in Studio tracks autonomous build agents tied to tasks.
Origin Agent in CDE
Origin Agent is the in-IDE assistant — it answers questions, edits files, and runs terminal commands when the task requires it. There is no separate mode to switch into; the agent is multifaceted on demand.
Typical workflow
- Open a cloud worktree or local project
- Open Origin Agent and describe the task — reference files with
@mentions when helpful - Review and approve tool calls as they execute
- Commit or push changes from the integrated Git UI
On cloud worktrees, tool calls (file edits, terminal commands) run inside the TDX sandbox — same runtime boundary as your code. Inference still routes through ORGN Gateway; see Origin Agent for ZDR vs TEE models.
Parallel agents via worktrees
Run multiple agents concurrently by giving each one its own worktree:
- Each worktree is a separate branch in a separate sandbox (or local tree)
- Agents on different worktrees do not stomp each other
- Create worktrees from the Projects sidebar per task
Recommended pattern: one worktree, one agent focus, one branch.
Agent Sessions in CDE desktop
Agent Sessions is a VS Code feature in CDE for browsing and managing long-running agent sessions — separate from the Origin Agent chat panel.
Where to find it
- Agent Sessions Welcome — optional startup editor tab with onboarding and session cards
- Agent Sessions sidebar — viewer in the chat auxiliary bar for filtering and refreshing sessions
Session providers
CDE labels sessions by where they run:
| Provider | Meaning |
|---|---|
| Local | Agent session on your machine |
| Origin CLI | Background agent running via Origin CLI in an isolated local worktree |
| Cloud | Cloud-hosted agent session |
Open the welcome page from the command palette: Open Agent Sessions Welcome.
Agent Session Projection
When viewing a session in projection mode, CDE focuses the workbench on a single session — title bar status, narrowed context, and an explicit exit path (Escape). Use this to review agent output without chat panel clutter, not to start new inference.
Agent chats in ORGN Studio
In the ORGN Studio browser app, the project sidebar shows Agent Chats when autonomous build agents are active on tasks.
This panel:
- Lists running and completed build agents per project
- Links each agent to its worktrees and browser Code Mode sessions
- Polls for status while agents are active
Studio agent chats are tied to the task board — they are not the same as Origin Agent in CDE or CDE's Agent Sessions window. Launching the desktop app from a Studio task is marked Soon; today, download CDE and sign in with the same id-orgn session.
Subagent models
Origin Agent doesn't do every part of a task itself: for well-defined sub-problems (searching a codebase, reasoning through a hard architectural question, reading a PDF) it delegates to a subagent, a focused helper that runs that piece and hands the result back. This keeps the main conversation efficient and lets each kind of sub-problem run on a model suited to it.
Settings → AI → Agents is where you choose which model each subagent uses. It's a routing page, not an authoring one, you can't create a new subagent here, only decide which model an existing one runs on.
By default every subagent is set to Inherits session model, it uses whatever model you picked for the current Origin Agent conversation, so switching your session's model changes every subagent's model too. Override a specific subagent when you want it pinned regardless of session, for example always sending oracle to your strongest reasoning model even in a session where you're otherwise using a fast, cheap model for everyday edits.
| Subagent | What it's for | Why you might pin its model |
|---|---|---|
general | General-purpose helper for multi-step work and parallel tasks | Keep on a fast, capable all-rounder for everyday delegation |
explore | Fast codebase search and navigation | Favor a cheap, fast model, this runs often and doesn't need deep reasoning |
oracle | Read-only deep reasoning for hard problems and architecture | Pin to your strongest reasoning model regardless of session default |
librarian | Looks up code, docs, and examples across repos and the web | A model good at synthesizing retrieved information |
multimodal-looker | Reads PDFs, images, and diagrams | Must be a model with vision/multimodal support |
There's also a separate Background Tasks entry, Small model, used for lightweight work like title generation, summaries, and conversation compaction, not a subagent invoked mid-task, and generally worth keeping on a small, fast, cheap model since the quality bar for these is low.
Tasks and automation
Studio tasks linked to worktrees can trigger automation when a worktree is created (runOn: worktreeCreated) — for example dependency install or environment validation before an agent starts.
Next steps
- Origin Agent — Gateway keys, models, and tool use
- Settings → Skills: reusable instructions agents pull in on demand
- Settings → Plugins: load global plugins for every project
- Cloud worktrees — create and switch worktrees
- ORGN Studio — task board and browser Code Mode
- Gateway quickstart — TEE models and inference receipts