Agent orchestration

Superconductor runs multiple agent sessions across separate workspaces, worktrees, terminals, and chats. A workspace gives you durable context, and Git-backed projects use task worktrees inside that context.

Task worktrees

A Git-backed task worktree has its own:

• Git worktree and branch.
• Terminal tabs.
• Chat sessions.
• Review diff.
• Target branch setting.
• Session history.

Review each worktree as its own branch.

Terminal-backed agents

Terminal-backed agents run as local processes in the selected project or worktree directory, so they use your shell, repo tools, credentials, and provider subscriptions. Terminal workflows include:

• Run Claude Code, Codex, Pi, Grok, Cursor Agent, Gemini CLI, Kiro, OpenCode, Copilot, or another CLI.
• Start a development server while another tab runs an agent.
• Run verification commands next to the conversation that requested them.
• Keep shell history and scrollback attached to the task.

Superconductor does not parse terminal output as an agent protocol. The terminal is a local PTY surface, and the provider CLI handles model traffic and authentication.

Chat-backed agents

Chat sessions provide a native conversation surface for supported providers. The local Superconductor app still owns them: the app owns the tab, project or worktree context, provider controller, local persistence, and event stream. Chat sessions expose:

• A native conversation interface.
• Model and reasoning controls exposed through the app.
• Local API automation with sc chat ....
• Session events that scripts can watch.

Worktree-aware prompts

Agents launched from Superconductor receive context that tells them how to inspect live worktree state. Worktree-aware automation should use:

sc worktree status --json

Use that command instead of guessing from the folder name or git defaults. Superconductor owns the worktree target branch, and the local API is the live source of truth while the app runs.

Coordinating parallel work

For parallel agent work:

1. Split the work into independent tasks.
2. Create a branch and worktree for each task inside the right workspace.
3. Start the relevant terminal or chat session inside each worktree.
4. Scope each agent to its worktree.
5. Review and merge each branch separately.

If two agents edit the same files, the resulting branches may conflict and need normal Git conflict resolution.

Automation hooks

The bundled sc CLI targets the current Superconductor session and worktree:

sc worktree diff-summary --json
sc worktree review-checklist --json
sc tab title "Review API Docs"

These commands let agents report the correct branch and review state without assuming details about the repository checkout.

CLI-driven agent teams

This is an experimental feature. Enable Settings → Experimental → Agent orchestration first; while it is off, the sc layout, sc agent(s), and sc coordination-state commands return a feature_disabled error. The manual multi-agent workflow above (separate worktrees, terminals, and chats) is not gated.

The sc CLI can also create and coordinate a team of agents directly, without manually opening each tab. sc layout run views|tabs|panes spawns labelled sessions across the layout and maps a prompt to each, sc layout send sends follow-ups to existing targets, and sc agents/sc agent/sc coordination-state group roles, broadcast prompts, wait for idle, and share machine-readable state:

sc layout run views --ui terminal --provider codex \
  --label implementer --prompt "Implement the change." \
  --label reviewer --prompt "Review the diff independently." \
  --output json
sc agent wait --to label:reviewer --idle --timeout-ms 120000 --output json

See Automation recipes and CLI and local API for the full command surface, selector grammar, and review-loop patterns.