Orchestration

Codius orchestration gives a coding agent control of the Codius daemon. The agent can discover every provider and model you have configured, create workspaces, launch other agents, send them follow-ups, and create heartbeats or schedules. The same work stays visible in the Codius app.

Native subagents vs Codius subagents

The most important difference from native subagents is that Codius subagents can cross provider boundaries.

Claude Code (Fable 5) => Codex (GPT-5.6)
Codex (GPT-5.6) => Grok Build
Cursor => Claude Code (Fable 5)

Native subagents belong to one provider. Claude Code launches Claude Code subagents; Codex launches Codex subagents. They are useful when the parent provider can handle the whole task itself.

Codius subagents are full agents managed by the Codius daemon. The orchestrator can choose any configured provider and model, keep the worker in the current workspace, or place it in another workspace created for the task. Use them when you want one model to plan, another to implement, and another to review.

Try it

Open Settings → your host → Agents, then turn on Enable Codius tools. Start a new agent, or reload an existing one so it receives the tools.

Then ask naturally:

Stay as the orchestrator. Use Codius to find my available Codex models, then
create a worktree-isolated workspace, then launch a GPT-5.6 subagent there. Ask
it to implement the parser change, run the focused tests, and report back here.

The orchestrator discovers the provider and model IDs, starts the worker, and receives a notification when it finishes. You can keep talking to the orchestrator in the meantime.

Agent creation has one default: when an agent creates another agent without a workspace ID, the new agent is its subagent in the same workspace. Passing a workspace ID changes where the subagent works, not who its parent is.

Where the work appears

Spawned work appears in the Subagents track above the composer. Open a row to read the live conversation.

Both kinds of subagent appear there:

• Codius subagents open as full agent sessions. You can talk to them directly, change their settings, or archive them.
• Native provider subagents open as read-only timelines. You can inspect their work, but their provider owns their lifecycle.

A cross-workspace subagent still belongs to its parent's Subagents track. Codius also opens its workspace so the work is not hidden in an otherwise empty workspace. If you want to turn any subagent into a top-level agent, detach it manually in the app or with codius agent detach; detachment is not an agent-creation mode.

If an agent says background work is running but the track is empty, update Codius. Provider-created subagent timelines require Codius 0.1.107 or newer.

Keep an agent working with a heartbeat

A heartbeat sends a prompt back into the same agent on a cron cadence. Use one when the agent should keep reassessing a live task: continue a refactor, babysit CI, watch a deployment, or retry after an external system changes.

Ask the agent directly:

Use Codius to create a heartbeat every 10 minutes. Keep checking this PR, fix any
new CI failures, and stop when all checks pass or after two hours.

The base /codius orchestration skill teaches agents how to create heartbeats, so you only need to ask. A heartbeat continues the current conversation; a schedule is better for standalone cron-style jobs such as daily triage.

You do not need to name MCP tools in your prompts. Ask for the workflow; the agent uses the tools underneath.

Continue with Common workflows for copyable prompts, Orchestration skills for packaged workflows, or the MCP reference for the complete tool catalog.