Disposable agents, durable memory: The architecture behind Squad (commandline.microsoft.com)

The development team behind Squad has unveiled a unique architecture that emphasizes disposable agent instances alongside a durable memory system integrated with Git for version control. Rather than focusing on making agents omniscient, the approach allows agents to spawn quickly and execute scoped tasks, preserving only the valuable output in a human-readable, inspectable format. This strategy not only facilitates collaboration among multiple agent teams but also enables rapid iteration and improvement through structured feedback mechanisms. A key testing ground, Squad Places, showcased the effectiveness of this architecture as different agent squads interactively posted comments and critiques, leading to immediate actionable commits. Within hours, teams identified and resolved various issues, demonstrating a seamless integration of agent-generated feedback into the development workflow. This adaptive system highlights the potential for effective multi-agent coordination without overcomplicating the agents themselves, ensuring that the human oversight remains integral while fostering continuous improvement. By prioritizing collective memory over individual agent persistence, Squad sets a new standard for collaborative AI development, suggesting a future where agent teams can evolve autonomously while still being accountable for their outputs.