This episode explores the rise of 'async agents'—background coding agents that run in the cloud autonomously. Walden Yan (Cognition, creator of Devin) and Cole Murray (OpenInspect) discuss the architectural shift from local to cloud agents, highlighting how model advancements have made fully autonomous coding practical. They cover key design decisions like harness in/out of the box, memory challenges, and the surprising reality that single-agent systems often outperform complex multi-agent setups.
Summarized by Podsumo
Devin's merge PR rate skyrocketed from 16% in January to 80% in March 2025, showing a dramatic shift to autonomous cloud agents.
OpenInspect chose to open-source its background agent system rather than commercialize, focusing on integrations and custom deployments for enterprises.
Both speakers agree that multi-agent systems are still overhyped; single agents with tool calls are more practical today, though 'smart pushback' from agents hints at future collaboration.
A key unsolved problem is 'memory'—how to maintain and retrieve contextual knowledge without exploding the agent's context window.
Common production use cases include SRE (first responder), PM-driven code changes via Slack, and customer support with full codebase context.
"We pretty much went from a specification to a completed pull request assuming the spec was good enough with very little friction. That paradigm changed a lot of how we interact with agents."
"_ Walden Yan"
"Having a background agent system is going to become critical infrastructure within a company. I open-sourced it so they can fork and customize."
"_ Cole Murray"
"Your codebase progresses to your worst engineer—if someone is gung-ho on AI and not auditing code, their patterns cement in, and the AI references them, exponentially growing the slop."
"_ Cole Murray"