Lou Bichard, Field CTO at Ona, presented on the topic of the “missing primitive” for agent swarms within the context of coding agents and software factories. He began by defining a software factory as a system that incrementally removes humans from the software development lifecycle (SDLC), aiming for automated workflows from development to production. Lou emphasized that while many are exploring this concept, the field is still in its early stages. He introduced different patterns for running coding agents at scale, such as swarms, fleets, and event-driven triggers, highlighting how these patterns help manage agents across multiple repositories and tasks.
Lou shared insights from his company, Owner, which has developed infrastructure to support these agent swarms, including the ability to spin up multiple development environments or virtual machines (VMs) to run agents in isolation. He demonstrated how agents can operate both within single VMs using process-based sub-agents and across fleets of VMs to handle large-scale tasks like CVE remediation or test coverage improvements. This approach addresses challenges related to security, resource contention, and scalability, with VMs providing stronger isolation compared to containers.
A significant challenge Lou identified is agent coordination—how agents interact, collaborate, and pass tasks among themselves. While runtimes, orchestration, and triggers are largely solved problems, coordination remains a gap. Existing tools like GitHub or Linear are not designed as coordination layers for autonomous agents, leading to noisy and overwhelming workflows for humans. Lou suggested that solving this requires breaking down the SDLC into micro-steps and managing context effectively, as context windows in large language models (LLMs) are limited and prone to degradation over time.
To address coordination, Lou proposed solutions involving state machines, workflow graphs, and durable executions that can represent the SDLC more granularly and deterministically. He mentioned prototypes using CLI-based tools that integrate with local or cloud agents to manage progression through SDLC steps. These tools could provide a standardized way for agents to check if tasks are complete and move forward, improving automation and reducing human intervention. Lou acknowledged that while some open-source projects and experimental frameworks exist, the field is still nascent and evolving.
In closing, Lou invited attendees to engage further through an upcoming virtual summit on software factories and background agents, as well as a public two-week project to build a software factory from scratch. He emphasized the importance of collaboration and sharing knowledge in this emerging area. During the Q&A, he elaborated on the coordination challenge and potential solutions, expressing interest in developing standards rather than focusing solely on implementations. Overall, Lou’s talk highlighted the promise and current limitations of agent swarms in automating software development and the critical need for improved coordination mechanisms.
