Swarm

Type /swarm to open the dashboard, then n to spawn one. Each agent is a separate zot subprocess with its own model loop, its own persistent session file, and its own chat in the dashboard — but they all run in the same working directory as the host, so they see and edit the same files you do.

There is no per-agent worktree or branch. If you need parallel edits on isolated checkouts, set that up yourself with git worktree outside zot.

Command line

/swarm                            # open the dashboard
/swarm new <task>                 # spawn an agent
/swarm new --model gpt-5 <task>    # pin the new agent to a model
/swarm logs <id>                  # jump into one transcript
/swarm send <id> <text>           # send a follow-up
/swarm resume                     # bring back a stopped agent
/swarm kill <id>                  # stop a running agent
/swarm remove <id>                # delete the agent's state

Continuous host and browser display

With Continuous execution, attached /swarm agents run as owned conversations on the host rather than local subprocesses. Their history lives in the continuous store rather than local Swarm session files. New children inherit their parent's current provider and model. Changing the parent later does not reconfigure existing children.

The browser client nests these conversations under their owner, including children of subagents. Parent arrows collapse or expand branches, and clicking a child opens its transcript. Search results remain flat, forks remain top-level, and children with a parent missing from the loaded list appear at the top level. This browser display does not change the TUI dashboard. See Continuous execution for browser setup and session settings.

Files and images

Task and follow-up editors support @ file references and !@ content selection. !@ sends UTF-8 text as file context and PNG, JPEG, GIF, or WebP files as images. Initial image tasks are sent through stdin rather than command-line arguments. Images persist in the agent session and are not resubmitted on resume. Use a vision-capable model for images. Restart agents started by older zot binaries before sending multiline or image follow-ups.

When attached to a continuous host, both @ and !@ transfer selected file content, so remote agents do not need the client's filesystem.

Persistence across restarts

Each agent keeps its event log and session under $ZOT_HOME/swarm/agents/<id>/. After restarting zot, previously spawned agents appear as detached. Open /swarm to inspect them or use /swarm resume <id> to continue one.

Startup reads only the last 4 MiB of each large event log. Opening the dashboard or resuming an agent loads its full log on demand, which can briefly delay that action for very large logs. The displayed transcript retains at most the most recent 2000 lines. The full event log remains on disk.

Auto-swarm

With /settings -> auto-swarm on, the main agent can fork sub-agents on its own via a built-in swarm_spawn tool when a request splits into independent parallel work. When the last sub-agent finishes, zot injects a single auto-swarm update message back into the main chat with each agent's status, task, and complete final response. If an agent produced no assistant response, the update includes a truncated transcript tail for diagnostics instead.