Skip to content

Project Index

Summary of all Orchestrator projects analyzed in this book

Detailed Project Profiles

Tmux-Orchestrator

Dimension Details
Language/Tech Stack Bash + tmux
Core Architecture Three-Layer Hierarchy (Orchestrator → Architect → Executor)
Communication Method bracket-paste injection + capture-pane monitoring
Fault Tolerance 2 layers (built-in heartbeat + systemd watchdog)
Knowledge Accumulation CLAUDE.md + LEARNINGS.md + FEATURES.md
Isolation Role separation (Architect manages plans, Executor writes code)
Deployment systemd user service + loginctl linger
Key Features Self-triggering agents, git discipline, cross-project coordination
Repository GitHub
Evolution Stage Stage 2 (reliable operation, progressing toward Stage 3)

Standout innovation: The bracket-paste protocol solves a real tmux problem — multi-line text corruption via send-keys. This is a "battle scar" pattern that can only come from production experience.

agency-agents-zh

Dimension Details
Language/Tech Stack Markdown + MCP Memory Server
Core Architecture Seven-Stage Pipeline (Recon → Plan → Review → Build → Test → Deploy → Monitor)
Communication Method MCP Memory (semantic search + rollback) + copy-paste handoff
Fault Tolerance 4 layers (quality gates at each pipeline stage)
Knowledge Accumulation MCP Memory (remember/recall/rollback)
Isolation Minimal (prompt-based role specification only)
Deployment install.sh script deploys to 10+ AI tool formats
Key Features 211 expert agents, DAG workflow, breakpoint resume, 32 handoff templates, cross-platform deployment
Repository GitHub
Evolution Stage Stage 1 (running, but lacks autonomous fault tolerance)

Standout innovation: Cross-platform agent deployment — a single agent definition can be installed as Claude Code, GitHub Copilot, Cursor, Aider, Windsurf, and more. This "write once, deploy everywhere" approach is unique in the ecosystem.

Composio (agent-orchestrator)

Dimension Details
Language/Tech Stack TypeScript + pnpm
Core Architecture Orchestrator-Worker (1:N with dashboard)
Communication Method Shared files (todo.md, scratchpad) + JSON handoff
Fault Tolerance 2 layers (LifecycleWorker + dashboard monitoring)
Knowledge Accumulation None (agents are stateless per session)
Isolation git worktree (Workers in independent worktrees)
Deployment ao start <url> one-click CLI
Key Features CI/CD integration, PR-based workflow, runtime agnostic, React dashboard, preflight checks
Repository GitHub
Evolution Stage Stage 2 (reliable, but knowledge gap limits Stage 3)

Standout innovation: Preflight mechanism — check all dependencies before starting, rather than discovering "git not installed" at runtime. This is a DevOps best practice applied to Agent orchestration.

Overstory

Dimension Details
Language/Tech Stack TypeScript + Bun runtime
Core Architecture Coordinator → Lead → Worker tree (hierarchical)
Communication Method SQLite Mail (9 protocol types + group addresses)
Fault Tolerance 4 layers (tiered watchdog: bash timer + AI triage)
Knowledge Accumulation Mulch Knowledge Base (conflict patterns, failure patterns, project knowledge)
Isolation git worktree (managed by WorktreeManager) + 4-level merge strategy
Deployment ov init + ov coordinator CLI
Key Features 11 runtime adapters, capability-based dispatch, overlay injection, event store, checkpoint handoff
Repository GitHub
Evolution Stage Stage 3 (intelligent enhancement, most mature)

Standout innovation: Overlay injection — a three-layer rendering system (role definition + project profile + task assignment) that ensures every agent starts with the right context. This is the "consumer side" of knowledge accumulation.

Reference Projects (Not Deeply Analyzed)

Project Language Architecture Key Contribution
Claude Code TypeScript Single-Agent CLI Native CLI integration, CLAUDE.md convention
CrewAI Python Role Collaboration Human-agent collaboration, shared memory
LangGraph Python State Graph Graph-based workflows, conditional edges, checkpoints
AutoGen Python Multi-Agent Conversation Conversational orchestration, code execution
OpenAI Swarm Python Agent Handoff Minimal handoff framework, lightweight

Cross-Project Comparison Matrix

Dimension Tmux-Orch agency-zh Composio Overstory
Setup Complexity Medium Low Low Medium
Max Agent Count 3-5 211 (catalog) 5-10 5-20
Autonomy Level High Low (human-driven) Medium High
Production Readiness Battle-tested Template-only Early stage Most complete
Cross-Runtime No Yes (10+ formats) Partial (4 runtimes) Yes (11 adapters)
Knowledge Persistence LEARNINGS.md MCP Memory None Mulch + Event Store
Merge Strategy None None Basic 4-level AI-assisted