WorkOS says factories must ship outcomes, not PRs

Ryan Cooke, an engineer at WorkOS, told AI Engineer that a real software factory is not a sandbox that spits out pull requests.

Ramp set the template in December 2025 when it published its Inspect architecture post, saying the agent wrote roughly 30% of merged frontend and backend pull requests.

WorkOS tried the standard recipe after that, running an OpenCode router on Cloudflare sandboxes to generate PRs, but Cooke said the results were indistinguishable from engineers just running Claude Code locally so the team shifted from counting PRs to measuring outcomes.

What followed was a two part factory, TARS living in Slack, Linear and GitHub and using webhooks to track projects and auto pick up the next blocked ticket, and Horizon orchestrating infrastructure in front of an internal MCP gateway Cooke calls a context engine.

That gateway connects to Snowflake and other internal tools, exposing semantic tables about product usage and customer conversations and letting agents or employees query data directly from Slack without bloating the context window.

What still has to happen, Cooke said, is a memory layer with evergreen context about people, teams and products, sandbox ownership to enable self improvement, and a credible answer for authorization that remains unsolved.