# Guy Podjarny: Building a Software Factory You Can Trust

## Executive summary

Guy Podjarny outlines the evolution of software development from 'building the car' to 'building the factory' using AI agents. He argues that building a trustworthy software factory requires a structured progression: defining reusable **skills**, orchestrating these skills within **loops** (agentic workflows), and finally aggregating them into a comprehensive **factory**. The talk addresses core concerns—Quality, Cost, and Safety—by detailing how skills can automate tasks like vulnerability remediation, how loops enable self-improvement, and how verifiers can manage agent empowerment and guardrails.

## Key takeaways

- The Shift to Software Factories: Software development is moving from manual, human-centric processes to automated systems where AI agents perform multiple stages (design, implement, review, test). The human role shifts to supervising and optimizing the factory itself. (00:00:00)
- The Factory Progression: Skills to Loops to Factory: To build a trustworthy factory, the process must evolve: 1) Define discrete, reusable **skills** (what good looks like); 2) Combine skills into bounded, self-improving **loops** (agentic workflows); 3) Aggregate loops into a comprehensive **factory** (business outcomes). (00:03:50)
- Automated Vulnerability Remediation: Skills can be used to automate complex tasks like vulnerability remediation. By defining a process (find, fix, verify, review, merge) and packaging each stage as a skill, agents can achieve high success rates. At Tesla, verified fixes showed success rates significantly higher than initial attempts. (00:07:47)
- Cost Reduction via Loops: Implementing self-improving loops can drastically reduce operational costs. Tesla reported reducing the cost per Pull Request (PR) from approximately $75 to $13 through optimized loops. (00:21:16)
- Safety and Empowerment Guardrails: Safety in a factory requires managing agent empowerment. This is achieved using **Tessel verifiers**, which are small instructions providing yes/no questions (e.g., 'Only act on PR feedback within the linked ticket's scope') to define strict boundaries and guardrails. (00:12:18)

## Technical details

- Software Factory Definition: A factory is the automation of the entire software development life cycle, typically manifesting as a 'ticket to PR' flow. This flow is powered by shared componentry (skills and tools) and must be self-improving. (00:06:24)
- Agentic Workflow (Loop): A loop is defined as a skill-powered agentic workflow, combining a skill, an execution environment (sandbox), and a trigger (user action or schedule). The output includes both the artifact (e.g., code) and telemetry (logs) for self-improvement. (00:17:11)
- Factory Components: A factory consists of three core areas: 1) A library of reusable components (skills, loops, infrastructure); 2) A specialized Operations UX for monitoring the factory floor; and 3) The ability to aggregate changes into defined business outcomes (e.g., ticket to mergeable PR). (00:24:03)

## Practical implications

- Organizations should view the adoption of AI agents not just as increasing productivity, but as increasing ambition and broadening scope (e.g., shifting GTM left).
- The goal is to move toward a state where developers build the factory that builds the product, rather than building the product directly.
- When evaluating third-party tools, consider the value of shared learnings and the ability of a partner to curate the technical world for you.

## Topics

Software Development Lifecycle, AI Agents, Automation, DevOps, Software Architecture, Tessel, AI DevCon London 2027

Source: https://www.youtube.com/watch?v=3NH_i8M1Gqg
