Feature

Swarm mode.

One leader chat, many follower agents. The leader delegates; you watch the vault work.

Mark any chat as a leader and it gains one new tool: spawn_agent. Ask for parallel work and the leader spawns follower agents — real, separate chats that each research, read, or draft one subtask with the full vault toolset, then report back. Every follower opens in its own pane (or window) so you can watch each one work.

Leaders are how you say "divide this up". Followers are how it gets done.

Making a leader

Either path, one step:

The pane's "..." menu → Make leader (spawn agents) — a crown appears by the title
The input box → /leader (repeat either to turn it off)

A leader chat always runs with tools enabled — no other setup. The setting Settings → Agent → Max follower agents (1–4, default 3) caps how many followers one message may spawn.

What a follower is

A real conversation — its own vault file, its own transcript, resumable and exportable like any chat. It runs on the leader's model, or a named specialist's own when spawned as one.
One task — the leader writes complete, standalone instructions; the follower can't see the leader's chat. It works with the vault tools (read/search/create/edit notes, web tools if enabled) and returns a full report.
Watchable — the run opens in a pane next to the leader. Each step lands as it happens. Close the pane whenever; the run keeps going and the transcript keeps writing.
Not a leader — followers cannot spawn further agents. One level of delegation, by design.

Skills on followers

Two ways skills ride a follower, both loading the same full instructions the skills system uses everywhere else:

Per agent — a named agent with pinned skills (set in the agent editor) brings them into every follower spawned as that specialist. The roster shows the pins (· skills: a, b) so the leader can match subtask to specialist.
Per spawn — the leader passes skills on spawn_agent for a one-off subtask ("spawn a follower with pdf-processing for this one"), whether or not a named agent is involved.

Followers also see the ordinary discovery index, so a follower can pull further skills itself via load_skill when vault tools are on.

Followers run one at a time — globally, across every leader chat. That keeps cost predictable and means two followers never edit the same note simultaneously. A follower spawned while another is still working queues behind it and starts when its turn comes. The leader collects each report, then synthesizes.

Stopping

Stop on the leader stops everything — the leader's response and every follower it started. Each follower's transcript records where it was stopped, including "before this agent started" for one still waiting in the queue. A follower's own pane is read-only while its leader owns it; the composer unlocks when the run ends. And while the leader is working, its own pane won't take a second send — one generation at a time per leader chat, so a second pane on the same leader can't tangle the transcript or reset the follower cap.

Where the transcripts live

Everything is ordinary history: leader chat, each follower chat, every tool call and report — searchable, exportable, and hand-editable in AI/Conversations. A follower's report is capped at what the leader receives (8,000 characters); the full text always stays in the follower's own transcript.

Swarm vs agent mode vs the arena

Agent mode — one chat using tools, one step at a time. The leader is agent mode plus delegation.
The arena — one prompt to 2–4 models, pick a winner. Comparison, not cooperation.
Swarm — a task divided into subtasks, each done by its own agent, results combined.

Good fits

Vault-wide research

"Find every note that mentions project Falcon, read them, and summarize the state of the project" — one follower reads, the leader synthesizes; or several followers split folders.

Drafting from sources

"Draft an outline from my interview notes" — a follower digs through the source notes while the leader holds the conversation and shapes the structure.

Parallel reading

"Compare these three papers' methodologies" — one follower per paper, reports come back side by side.

Deliberate limits (v1)

No ad-hoc per-follower model — a follower runs on the leader's model, or a named specialist's own via the agent param; picking an arbitrary model per spawn (without creating a named agent) is not offered.
No follower-to-follower messaging — the leader coordinates; followers report to it. This is the shape of multi-agent work that actually holds up: isolated, read-heavy subtasks, one synthesizer.
Serial execution — one follower runs at a time across all leaders; extra spawns queue behind it and start in order.
Shared vault — followers can create and edit notes like any agent chat. Leader-delegated edits are serialized (one follower at a time), but a follower and your own edits can still meet in the same note — same as two panes today.