The Missing Primitive for Agent Swarms
summary
Claim: runtime and orchestration are largely solved but a coordination primitive is missing. Examples: Stripe called theirs Minions and RAMP called theirs Inspect are homegrown fleets that drive thousands of parallel pull requests, Owner demos VM‑level isolation and sub‑agents that spawn fleets of VMs (scale limited only by cloud), and GitHub proves noisy and unusable as a coordination layer. Proposed concrete solution is a CLI gateway that agents invoke to check state, pass work, and gate progression.
AI-written summary · about
chapters (11)
- 0:00Introduction and definition of a Software Factory
- 1:50Agent swarm patterns: Swarms, Fleets, and Events
- 3:11Real-world examples of internal agent infrastructure (Stripe, RAMP)
- 3:50How Owner handles agent infrastructure and development environments
- 4:49Understanding Harness Engineering
- 5:43The three pillars of agent swarm infrastructure: Runtime, Orchestration, and Coordination
- 7:17Demo: Running sub-agents and fleets in Owner
- 10:20Challenges of building a software factory
- 11:44The issue with Context Management and Context Rot
- 12:16Why GitHub is a poor coordination layer for agents
- 12:59Proposed solutions: State machines, Durable execution, and CLI gateways
related talks
25mBeyond the IDE: Toward Multi-Agent Orchestration
24mAI Agent Swarms Are Amazing
31mThe Future Is Domain-Specific Agents
15mAgents Don’t Ship Themselves
27mWhy the Best AI Agents Are Built Without Frameworks (Primitives over Frameworks)
21mDemystifying AI Agents: A New Primitive for Flow Control
25mBeyond Orchestration: An Open Source Architecture for Agentic Systems
27mAgentic Systems: Leadership, Risk, and Accountability
