Multi-Agent Orchestration
The practice of coordinating multiple AI agents working in parallel, each with distinct roles and tools, toward shared organizational goals. A step beyond single-agent use toward persistent, autonomous AI “companies.”
The Core Problem
A single Claude Code session handles one thread of work. As usage scales — multiple projects, multiple workflows, background tasks — you end up with many parallel sessions that each require attention, lack shared context, and have no unified visibility. The “20 terminals” problem: which session is doing what? What did you ask each? What needs approval?
Multi-agent orchestration solves this with centralized dashboards, ticketing systems, and role-based agents that operate autonomously and surface work to the human only when needed.
Key Concepts
Agent Roles
Agents are not generic — they are configured with specific personas, tools, and behavioral constraints. Common org-chart structure: CEO (strategic, sets goals, makes hires), Engineers (implement), QA (review and test). Each role has different permissions, tools, and reporting relationships.
Heartbeats
Agents wake on a schedule with fresh context, review their task queue, and continue work. They don’t need a human to activate them — they operate around the clock. On wake, agents must re-orient (check instructions, review open tasks) before acting, since they have no persistent memory between sessions.
Human-as-Board Pattern
The human’s role shifts from operator to board member. You set high-level goals and metrics, approve major decisions (new hires, large scope changes), and review results — but you don’t manage individual tasks or sessions. Nate Herk’s framing: “30 minutes as the board” rather than hours in the terminals.
Inbox and Approvals
Requests that require human judgment surface as inbox items (approvals, questions, blockers). Everything else runs autonomously. Board approval can be toggled off per-category as trust in the system grows.
Agent Memory via Files
Agents have no cross-session memory by default. Their “memory” is maintained through:
- Task history in the ticketing system
- Soul/heartbeat/agents files that persist between wakes
- Project documentation and CLAUDE.md files
Relation to Agentic Harness Primitives
Agentic Harness Primitives defined three permission handler tiers: interactive (human in loop), coordinator (orchestrator handing out permissions), and swarm worker (autonomous under orchestrator). Multi-agent orchestration tools like Paperclip implement all three layers.
The 12 primitives become more critical at multi-agent scale: workflow state prevents duplicate work across agents; permission audit trails become essential when many agents can take real-world actions; agent type systems (Explore, Plan, Verify) prevent agents from overstepping their roles.
Tools
- Paperclip — open-source, MIT, dashboard-based; agents run as Claude Code sessions; 36k stars
- OpenClaw — the earlier tool this pattern popularized (native heartbeats)