If your agents are doing the work, the network effects that make or break the product are being seeded right now — and most founders are too busy watching outputs to notice they're engineering a cold start into their architecture.
Andrew Chen's The Cold Start Problem reframes growth as a problem of network density before scale. The insight that sticks is the atomic network: the smallest cluster of users and supply that can sustain itself without outside help. Chen is specific that you don't fight the cold start by acquiring more users — you fight it by identifying the hard side, the scarce, difficult-to-recruit participants whose presence makes the network worth joining for everyone else, and solving their problem first, often manually, at a loss, with no elegant automation in sight.
The hard side doesn't disappear when agents replace human labor — it migrates. In an AI-native product, the hard side is usually the data, the domain context, or the trust signal that makes agent output worth acting on. Your atomic network is the smallest set of workflows, human reviewers, and feedback loops that can validate agent judgment without collapsing into noise. Founders who skip this step and scale agent volume early are just manufacturing cold starts at speed. Chen's framework says: find that minimum viable trust unit, make it work ugly, then grow it.
- Identify your hard side — it's probably the human who validates agent output, not the agent itself
- design your atomic network as a trust loop before you design for throughput
- escape velocity is earned by density, not by adding more agents to a system that hasn't proven it can sustain itself yet.
