How to Build an AI-Native Software Factory

Summary

This article argues that companies do not need to build their own coding agents to create an AI-native software factory. Instead, they should build the shared system around purchased agents: cloud sandboxes, model and tool gateways, reusable skills, context, workflow harnesses, identity, and human approval gates. Drawing heavily on Uber’s published engineering work, it explains why laptop-based sessions fail to support unattended workloads, consistent company-wide practices, cost visibility, and review capacity as output rises. Uber organizes its agents into specialized workers, a general front door, developer sessions with shared skills, and raw coding sessions; other examples include Shopify’s River, Spotify’s maintenance agent, Stripe’s Minions, and DoorDash’s review workflows. The recommended starting point is measurable toil such as migrations or code review, run in shadow mode with a purchased hosted agent. The article stresses that deterministic automation should replace an agent when a task can be expressed as a reliable rule, citing Uber’s Shepherd migration of more than 75,000 test classes. It describes six platform blocks: a model gateway, isolated development environments, a tool catalog or MCP gateway, shared context, a harness with triggers and workflows, and security gates. Those gates include scoped agent identities, logging, human review before production, and controls such as Meta’s Rule of Two. For rollout, teams should measure review queues from Git history, track cost per merged outcome alongside quality and volume, and keep developers involved through opt-in adoption, visible results, and human accountability. The central buy-versus-build recommendation is to buy agents and commodity infrastructure while building the company-specific code snapshot, tool catalog, context, workflows, and policies. The article notes that its public examples come mostly from large technology companies, so smaller organizations should treat the sequence as a starting point and expand only when each workflow proves its value.