Skip to content

Orchestrating Agents: From Hard Guards to Soft Constraints

A systematic guide to AI Agent orchestration — when agents run 24/7, who watches the watchers?

Star on GitHub

What This Book Covers

This book systematically explores how to orchestrate AI Agents for autonomous, long-running work. It draws from real-world projects and distills them into reusable patterns.

Research Foundation: Four Production Systems

This book analyzes four production-grade orchestrator systems, each addressing orchestration challenges in unique ways:

Project Architecture Key Innovation Best For
Tmux-Orchestrator Terminal-based send-keys + capture-pane monitoring Simple agent coordination
agency-agents-zh Human-in-loop Swarm handoff templates + MCP memory Creative workflows with oversight
Composio File-based Shared todo.md coordination Minimalist team coordination
Overstory SQLite mail Typed protocol messaging Large-scale automation

Key Research Findings

Communication Patterns Analysis

The research reveals a clear evolution from simple terminal interaction to structured messaging: - Bracket-paste: Most reliable for multi-line commands - SQLite mail: Best for machine-to-machine coordination - Swarm handoff: Optimal for human-AI collaboration - Shared files: Simplest for small teams - MCP memory: Most flexible for cross-session context

Architectural Patterns

  • Hierarchical delegation: Orchestrator → Team Lead → Specialist Workers
  • Session isolation: Git worktrees prevent cross-contamination
  • Fault tolerance: Multiple recovery strategies (restart, rollback, escalate)
  • Quality gates: Automated validation before task completion

Implementation Insights

  • Type safety: Strong typing prevents communication ambiguity
  • Persistence: Messages must survive agent crashes
  • Explicit routing: Don't make agents "guess" communication paths
  • Group addressing: Essential for scaling beyond 3 agents

Part I: Hard Orchestration — Low-level hard logic that keeps agents running:

Chapter Topic Key Content
1 Introduction Why orchestration, five core pain points
2 Architecture Patterns Five topologies from dual-agent to tree
3 Role Systems Who does what, persistent vs ephemeral roles
4 Communication send-keys, bracket-paste, SQLite mail
5 Fault Tolerance Watchdog, progressive recovery, fast-crash protection
6 Isolation & Concurrency git worktree, process isolation, file locks
7 Deployment & Daemons systemd, tmux, nohup, cron
8 Rule Guard Iron law blocks, external guard scripts, integrity checks

Part II: Soft Orchestration — Skills and prompts that shape agent behavior (agents read these):

Chapter Topic Key Content
9 Prompt Engineering Iron laws, mission injection, SPRINT-driven
10 Skill Systems Reusable skills, SKILL.md, templates, MCP tools
11 Knowledge Accumulation LEARNINGS.md, MCP memory, experience docs
12 Pipeline Orchestration Multi-step workflows, quality gates

Part III: Practice & Evolution — Build and evolve your own orchestrator:

Chapter Topic Key Content
13 Antipatterns Must-avoid pitfalls across hard and soft orchestration
14 Hands-On Build a minimal orchestrator from scratch
15 Evolution Roadmap From scripts to autonomous systems

Hard vs Soft: The Core Distinction

Hard orchestration = low-level hard logic. It runs as infrastructure (daemons, scripts, cron jobs) that agents cannot modify. It enforces constraints externally — if an agent crashes, the watchdog restarts it; if an agent deletes its rules, the guard script restores them.

Soft orchestration = written in skills and prompts for agents to read. It shapes behavior by providing instructions, experience, and reusable templates. It depends on the agent's compliance — the agent reads and follows, but could theoretically ignore it.

The most robust systems combine both: hard orchestration enforces the non-negotiables, soft orchestration guides everything else.