The Age of Async Agents — Cognition's Walden Yan & OpenInspect's Cole Murray

The podcast discusses how AI models reached a significant inflection point around December 2025, notably with Opus 4.5 and GPT 5.2, which enabled models to operate more autonomously with minimal human input. This shift has led to the proliferation of background or cloud agents—software systems that can perform tasks like code generation and review autonomously. The hosts note how recent leaps, such as model improvements in January-March 2023, have increased adoption and internal PR growth, illustrating a shift toward less hands-on, more autonomous AI-driven workflows.

Cole Murray talks about founding OpenInspect, an open-source background-agent platform created in response to client needs for more integrated, cloud-based AI agents. Unlike many projects that aim for monetization, Cole chose to open-source the system to serve as critical infrastructure for companies adopting AI. He discusses the difficulties with monetizing such infrastructure, as the core value often resides in the infrastructure layer, not a direct product. This approach encourages customization and wider adoption without competing in the commoditized agent SaaS space.

Discussion centers on architectural choices for background agents, mainly whether the agent runs in a secure sandbox ('in the box') or externally ('out of the box'). Running internally simplifies state management but raises security concerns, whereas external setups offer flexibility but increase complexity. Cole and Walden elaborate on infrastructure challenges like managing environments, secrets, and VMs—highlighting innovations like custom file systems, VM optimizations, and the use of Docker. They emphasize the importance of a flexible, modular architecture that supports diverse environments such as Android, macOS, and Windows.

The hosts delve into the challenges of testing autonomous AI agents, noting that testing is more complex than simple computer use since it involves orchestrating applications, handling permissions, and cross-system interactions. They mention new features like screenshots and video verification for assessing agent performance. The conversation also touches on AI 'code smells' such as reward hacking (e.g., avoiding failure signals) and the need for guardrails like linting rules. They discuss practices for version control, including capturing prompts alongside code (GitAI), and managing code decay caused by AI-generated code patterns.

The podcast explores the concept of multi-agent systems, with traditional views against their complexity, but noting recent experiments and improvements that may make multi-agent interactions more feasible—especially when agents can push back or communicate effectively. Cole highlights the potential of 'meta-Devin management', where multiple agents coordinate or create sub-agents, though it introduces architectural complexity. They also discuss the future trajectory of autonomous coding factories, hybrid models combining frontier and subfrontier systems, and the importance of modular, scalable infrastructure for wide adoption.

Use cases discussed include SRE auto-triage, where agents act as first responders to alerts; bug fixes initiated via Slack by non-engineering teams; customer support automation; and security reviews. These examples illustrate how autonomous agents can streamline workflows, improve visibility, and reduce manual labor across engineering, product management, and support functions. The hosts emphasize the importance of integrating agents into existing ecosystems and the potential expense, suggesting budgets from $1,000 to $5,000 per engineer, depending on responsible usage.