No, That's Not a Software Factory — Ryan Cooke, WorkOS
This post critiques the standard 'software factory' model where an agent is placed in a sandbox and merges PRs, arguing that this approach yields results indistinguishable from engineers using local AI tools. The author proposes an alternative 'orchestra' model where the engineering process itself is built into the factory. Key contributions include: 1. Redefining success metrics from PR count to customer-facing outcomes like defect rates and recovery time. 2. Introducing TARS, an agent that integrates with Slack, Linear, and GitHub via webhooks to manage project tickets. 3. Describing Horizon as an orchestration layer that connects to an internal MCP gateway, which acts as a 'context engine' for all tools, including Snowflake. 4. Highlighting the importance of a PM agent for drafting product docs ('Hilltop') and breaking down approved work into tickets, thus solving the 'blank page' problem. 5. Discussing the need for custom sandboxes and a company-wide memory layer for enhanced agent capabilities and self-improvement.
More from the Software Factory Engineering trail
- From Your Laptop to the Pipeline: Scaling Custom Agents with GitHub Copilot
- Generation Is Cheap, Review Is Expensive: How to Stop Shipping AI Slop — Gabriel Martinez, G2i
- Let Anyone at Your Company Ship Internal Apps with AI — Garrett Galow, WorkOS
- The Software Factory: From Bug Report to Production Code — Davis Palmie, Factory
- Why AI Agents Should Have Their Own Sandbox — Philipp Schmid, Google DeepMind