WorkOS Reimagines Software Factory Goals

This standard setup focuses on automated code generation and merging, representing a common industry approach to leveraging AI in development workflows.

Focusing solely on the number of pull requests or lines of code generated by AI agents can obscure the actual performance and effectiveness of a software factory.

An increase in PR volume does not automatically translate to improved customer outcomes, prompting WorkOS to prioritize outcome-based metrics over simple generation volume.

By providing each engineer with a small, dedicated engineering team, WorkOS aims to multiply individual capacity and enable the shipping of more complex features at higher velocity.

The primary measure of success for such a factory is the ability to ship more features faster, focusing on building more complex functionalities with increased velocity rather than just generating code.

The aim is to enhance overall productivity and accelerate product development cycles.

The system architecture incorporates TARS, an interaction layer embedded within Slack, Linear, and GitHub, alongside Horizon, an infrastructure orchestration layer that utilizes an MCP gateway.

TARS uses webhooks to track project progress beyond mere code generation, while Horizon provides the underlying infrastructure for autonomous product engineering, moving beyond simple code generation to actual product engineering autonomy.

Webhooks allow TARS to track Linear tickets, identify dependencies or blocking issues, and automatically initiate the next cycle task once a preceding ticket reaches completion.

This agent-driven system re-evaluates and updates project plans in real-time as gaps are identified, enabling autonomous execution of predefined project maps.

This approach streamlines the workflow and ensures continuous progress without constant human intervention for task assignment.

A 'hilltop document,' functioning as a Product Requirements Document (PRD), defines project purpose, incorporates customer feedback, competitive analysis, and design mockups, standardizing information gathering across the company.

This agent breaks down project specifications into manageable tickets, while humans retain control to refine, edit, and provide guidance on the AI-generated outputs.

This system provides a structured framework for the lead product engineer to refine specifications and automate project management tasks, allowing engineers to dedicate more time to high-level implementation.

The integration streamlines project setup and resource allocation, enhancing efficiency in the initial phases of development.

WorkOS's system allows for the integration of various coding agents, such as Devin or Claude Code (Opus), rather than being restricted to a single one.

The Model Context Protocol (MCP) enables these diverse agents to pull relevant context from internal systems and project documentation, ensuring they have the necessary information to implement different parts of the work.

Cross-functional teams, including security, participate directly in project channels, providing oversight and collaboration within this flexible development environment.

This engine manages system prompts and provides instructions for tool navigation, connecting directly to data lakes like Snowflake to supply semantic context for AI agents.

The Context Engine ensures agents can effectively interact with various internal systems and understand the nuances of organizational data.

This centralized access allows for Slack-based data queries for customer analysis, making organizational data readily available to AI agents.

WorkOS recommends this as a foundational step for any organization building a software factory, emphasizing the importance of an MCP gateway for effective tool integration and agent guidance.

WorkOS is developing its own sandbox infrastructure to gain deep control over session data, allowing workloads to be moved across different parts of their infrastructure.

This self-sufficiency in infrastructure management provides greater flexibility and control over their development environment.

This memory layer aims to enable interoperability between the software factory and other AI tools, centralizing organizational knowledge.

The goal is to provide a comprehensive, up-to-date semantic understanding of WorkOS operations.

WorkOS prioritizes delivering customer value and continuously improving its infrastructure, moving beyond simple output metrics.

The company tracks metrics such as defect rates and time to recovery, and uses infrastructure data to identify skill gaps and evolve the software factory.

This approach aims to keep pace with rapid AI developments by constantly refining techniques and ensuring relevance.

WorkOS asserts that the true value of a software factory lies in its ability to deliver tangible value to customers, rather than merely focusing on volume metrics like PR counts.

The factory aims to embed customer impact directly into its automated processes, with engineers closely monitoring defect rates and time to recovery as critical quality indicators.

Increased usage of cloud sandboxes over local harnesses is viewed as a sign of genuine engineering productivity, reflecting the system's effectiveness in real-world application.

Owning its infrastructure enables WorkOS to monitor the factory's performance, identify areas for improvement, and track where AI agents make mistakes.

This data helps determine which new skills should be developed for the agents and if older skills have become obsolete, ensuring the factory remains current.

The system aims to keep pace with rapid AI developments by continuously identifying and integrating relevant techniques, thereby ensuring the factory's ongoing evolution.

While sandboxes and AI-generated pull requests are useful, WorkOS prioritizes embedding comprehensive software engineering practices and processes into automation.

Ryan Cooke invites other developers to collaborate on unresolved challenges, particularly authorization within software factories, and maintains a booth at the event for discussions on factory architecture.