Swarms & formations — Conductor V2

The formations

Three shapes. The formation decides who talks to whom — and who starts.

🏛

Hierarchical

Orchestrator → workers

      ┌─ ORC ─┐
  ┌───┼───┬───┼───┐
  w1  w2  w3  w4

One orchestrator decomposes the mission, delegates a single task per worker, collects their reports, and synthesizes the final result.

Best when the mission splits into independent chunks.

⛓

Pipeline

Sequential stages

s1 ─▶ s2 ─▶ s3 ─▶ s4
 each stage hands
 off to the next

Stages run in order; each consumes the previous stage's output and hands off to the next. Initiator is the first stage.

Best when there's a natural assembly line — recon → audit → verify → report.

🕸

Mesh

Peer-to-peer

  p1 ─── p2
   │ ╲ ╱ │
   │ ╱ ╲ │
  p3 ─── p4

Equal peers self-organize: each claims a distinct angle, works it, and broadcasts findings to the others. Peer 1 doubles as scribe.

Best when you want breadth — sweep a space from several directions at once.

How they coordinate

Two dumb, reliable channels

No message bus, no new infrastructure. A swarm coordinates the way people do: shared files and short pings.

→ Files — a shared swarm directory ~/.conductor2/swarms/<name>/ holds the mission, out/ artifacts, and notes/. Files are the source of truth.
→ Messages — swarm-say <window> "<one line>" for handoffs and pings; anything longer goes in a file with a pointer.
→ Each window's role briefing is written to disk; the window is kicked off with a one-liner pointing at it — long prompts never ride through tmux send-keys.
→ A baked-in member allowlist means a window can't message outside its own swarm; launch order is receivers-before-initiators so no ping lands on a window that doesn't exist yet.

~/.conductor2/swarms/x402-research/

# the swarm's shared brain
mission.md
prompts/      # per-agent briefings
  p1.md  p2.md  p3.md  p4.md
out/          # real work products
  landscape.md  figures.csv  REPORT.md
notes/        # scratch
swarm-say     # member-scoped ping helper

Presets

Premade swarms — fire as-is, or edit the purpose first

🔬

Deep Research

hierarchical · 5 agents

Orchestrator splits a research question across 4 researchers — landscape, evidence, red-team (counter-evidence), numbers — then synthesizes one cited, decision-grade report.

📈

Market Bot Management

mesh · 3 agents

Three peers each take one trading desk: confirm the bots are alive, audit the last 24h from the ledgers, and propose tuning. Read-only — changes go in the report for a human.

🛡

Web3 & Security

pipeline · 4 agents

Sequential review: recon maps the attack surface, the auditor digs line-by-line, the verifier reproduces or kills each finding, the reporter writes a severity-ranked fix list.

The swarm MCP

Plan and fire a fleet from any agent

Conductor V2 speaks MCP over stdio, so any MCP-aware agent can design a swarm, preview it, fire it, and stop it — natively. Read tools are free; the two that spawn or kill processes are flagged.

Tool

What it does

Kind

list_formations

The three formations with diagram + when-to-use.

read

list_presets

Premade configs (deep-research, market-bots, web3-security).

read

plan_swarm

Dry run — resolve a config into concrete agents + briefings, launch nothing.

read

list_swarms

Live swarms grouped from the managed-window registry.

read

fire_swarm

Launch the fleet — N tmux windows, each kicked off with its briefing. All-or-nothing.

spawns

stop_swarm

Kill every window of a swarm. Irreversible — confirm first.

kills

Watch a fired swarm from the V1 cockpit too — every window is a normal Claude Code session, so the claude-code adapter sees them, grouped by status, problems first.

Get started

Fire your first swarm

Clone V2, register the MCP, and ask any session to plan and fire a fleet.

# clone V2
$ git clone https://github.com/yksanjo/conductor2 ~/conductor-v2

# register the swarm MCP (user scope = everywhere)
$ claude mcp add conductor2 --scope user -- node ~/conductor-v2/mcp.js

# then, from any session:
› plan a pipeline swarm of 4 to security-review ~/soag-gate,
  show me the briefings, then fire it

Conductor V2 on GitHub → ← Back to Conductor V1