Eric Zakariasson from Cursor shares his insights on building a software factory using AI agents to automate software development. He outlines the progression of autonomy levels in software creation, from basic autocomplete to fully autonomous “dark factories” where agents handle coding, testing, and deployment with minimal human intervention. Eric emphasizes the benefits of such factories, including increased throughput, consistent output, and enhanced creativity, while acknowledging the complexity and effort required to build and manage them effectively.
To build a software factory, Eric highlights the importance of establishing clear primitives and patterns in the codebase, such as modularization and consistent usage patterns, which help agents navigate and work autonomously. Guardrails like rules, hooks, and tests are crucial to keep agents aligned and prevent costly mistakes, especially in sensitive areas like authentication or encryption. He also discusses enabling agents with skills and access to external contexts, such as feature flagging and dev environments, to empower them to operate independently and scale efficiently.
Eric demonstrates Cursor’s approach with their new Cursor 3 platform and cloud agents that run in isolated virtual machines, allowing parallel work without interference. He showcases practical examples like automated end-to-end testing using Playwright and agents that can verify their own work by interacting with the UI, recording sessions, and performing code reviews. This automation reduces the need for humans to constantly monitor code, shifting their role from workers to managers who oversee and guide multiple agents asynchronously.
Managing a software factory requires a shift in mindset towards asynchronous workflows, trust in agents, and effective scoping to avoid conflicts. Eric stresses the importance of preserving tribal knowledge, maintaining clear intent, and dynamically evolving rules based on agent behavior. He also discusses the challenges of collaboration within teams when creating shared rules and guardrails, emphasizing the need for cultural alignment and forums for discussion to unify efforts and avoid siloed agent setups.
In the Q&A, Eric addresses concerns about code quality, security, and architectural integrity, suggesting a combination of human oversight, automated testing, and continuous refinement of rules. He shares insights on integrating issue tracking with agents, running agents locally or in the cloud, and balancing prototype development with production-ready code. Eric envisions future engineering roles evolving towards managing agent fleets and orchestrating workflows, highlighting the ongoing need for human creativity, accountability, and strategic planning in an increasingly agent-driven development environment.
