swarm
liveFire-and-forget multi-agent orchestrator
Type what you want in natural language. The LLM parses your intent, spawns parallel agents,
streams live observability, and writes immutable artifacts to timestamped run directories.
One command. Multiple agents. Reports when done.
swarm "explore architecture on creds-manager"
swarm "2 agents on portal-api find security issues"
swarm # interactive mode — pick repo, task, model, agent count- LLM-powered intent parsing — no flags to memorize
- Interactive CLI with repo picker and model selection
- Parallel agent execution via Claude Agent SDK
- Live observability — see every READ, GREP, WRITE in real-time
- Immutable run artifacts in _agent-runs/
/ - Conductor agent synthesizes multi-agent findings
- Per-agent cost and timing breakdown
/spawn-agent
liveSpawn any LLM agent on any repo in tmux
Create an agent on a repo with one command. Supports Claude, Grok, Kimi, and Codex.
Each agent launches in its own tmux session with native prompt support — the task
starts executing immediately, no manual typing needed.
/spawn-agent creds-manager cgod to explore architecture
/spawn-agent portal-api grok find security issues
/spawn-agent ~/repos/myproject kimi review CI/CD- Multi-LLM support — Claude, Grok, Kimi, Codex
- Native prompt injection per agent (positional args, -p flag)
- Tmux session management — new session, add window, detect conflicts
- Fuzzy repo resolution from registry
- Natural language task parsing
/agent-swarm-tmux
liveVisual swarm — 4 agents in split tmux panes
The visual version of swarm. Splits one tmux window into 2-6 panes,
each running a Claude agent with a different role. Attach and watch
all agents working simultaneously. Pure chaos, pure fun.
/agent-swarm-tmux portal-api architecture, security, ci/cd, bugs
/agent-swarm-tmux creds-manager code quality, test coverage- 2-6 split panes with tiled layout
- Each agent gets a specialized role prompt
- Reports written to _agent-reports/
.md - Visual — watch agents explore in real-time
Orchestration mental model
LLM as brain, not wrapper
An orchestrator that just launches agents with fixed roles is a wrapper, not an orchestrator.
Real orchestration means the LLM is the brain — it parses intent, decides what agents to spawn
based on what the repo actually contains, reacts to findings, and synthesizes with judgment.
Agents mental model
Agents are deployed workers, not chat assistants
Agents are workers you deploy, not assistants you chat with. Once launched, you don't re-engage —
you come back to read what they wrote. The conversation is observability, not the product.
Artifacts ARE the output.
Meta-engineering
Building the system that builds the system
The highest-leverage work is building the system that builds the system.
The tool is the product, not the output of the tool.
Don't just spawn an agent — build a system to spawn agents.
Don't just orchestrate — build an orchestration framework.