Everything Gets Rebuilt: The New AI Agent Stack | Harrison Chase, LangChain

The podcast discusses how AI agents have transformed from simple prompt-based systems into complex, multifunctional entities capable of planning, tool usage, code writing, and memory management. Harrison Chase emphasizes that the real advancement lies in the surrounding infrastructure—harnesses, subagents, and systems that enable these models to operate effectively in real-world applications. He highlights the importance of primitives like planning tools, system prompts, and context management that have become foundational to modern AI agent architectures.

Chase describes the key primitives that constitute the current state of AI agents: harnesses, subagents, file systems, and compute sandboxes. Harnesses act as the control layer, managing how models interact with tools and environments. Subagents facilitate context isolation and task management, while the file system enables long-term context handling and knowledge persistence. Sandboxes are critical for safe code execution, isolating untrusted code, and ensuring security, especially in cloud environments. These components are modular, flexible, and designed to support scalable, reliable AI agent deployment.

The podcast covers recent improvements in models and their impact on agent design, noting that better models have made loop-based execution and tool calling more reliable. Harrison emphasizes that the 'harness'—the system around the model—is crucial for performance, reliability, and security. He shares insights into practical implementations like LangChain, LangSmith, and Deep Agents, each tailored for different stages and use cases. The discussion also touches on long-term memory, context compression, evaluation, and continuous learning, underscoring a trend toward more autonomous, adaptable, and observable systems.

Chase underscores the importance of sandboxing for secure execution of untrusted code and management of sensitive information like API keys. Daytona is highlighted as a leader in providing sandbox environments that prevent prompt injection attacks and enable safe code execution in the cloud. The conversation highlights various ways sandboxes are used—either hosting agents locally or calling external sandboxes as tools—and discusses security considerations, such as protecting API keys from theft or misuse.

Looking ahead, Harrison discusses the ongoing evolution of agent infrastructure, including deploying long-running, stateful agents and building comprehensive platforms for agent management. He emphasizes that the main differentiation will shift toward instructions, tools, and domain-specific knowledge encoded as skills rather than raw model power. The focus is on creating robust, scalable systems that can adapt, learn, and improve continuously. He also mentions that they recently raised $125 million to advance these initiatives, with a strong emphasis on observability, deployment, and a holistic approach to agent engineering.