Uploaded November 2026 | Updated November 2026, 8 hours ago
Erik Bernhardsson, CEO and Co-Founder of Modal, explains what drove demand for Modal's sandboxes (vibe coding apps, background agents, and reinforcement learning rollouts) and why he thinks agent harnesses should run outside the sandbox, with sandboxes that are short-lived, maybe one per tool call. He also covers how Modal, which builds AI infrastructure for inference, training, and sandboxes, moved up the stack with auto and shared endpoints for open source models, when he thinks RL is worth its cost over supervised fine-tuning, and the multi-node training and discounted batch products Modal is building. The conversation closes on how Modal's engineers use coding agents and why Erik is watching voice to voice models.
This session was captured at Interrupt New York City, the agent conference by LangChain.
Chapters:
0:00 What Modal builds for inference, training, and sandboxes
0:57 Was Modal built for AI workloads before ChatGPT?
2:47 What made Modal's sandboxes finally take off?
5:35 Why Modal is moving sandboxes from gVisor to VMs
7:11 Should the agent harness run inside the sandbox?
10:09 Do agents need a full VM or a lightweight sandbox?
12:27 Are open source models close enough to swap in?
15:49 When is RL worth it over supervised fine-tuning?
17:08 Why Modal built auto and shared inference endpoints
18:43 Can reinforcement learning be packaged as a product?
19:44 What Modal is building next for training and batch jobs
21:09 Why Modal built a team of forward deployed engineers
23:25 How Modal uses coding agents and MCP servers internally
26:41 Why Erik is bullish on voice to voice models
Resources:
→ Modal: modal.com
→ Modal Sandboxes: modal.com/products/sandboxes
→ Deep Agents with Modal sandboxes: docs.langchain.com/oss/python/integrations/sandboxes/modal
→ Interrupt, the agent conference by LangChain: interrupt.langchain.com
Erik Bernhardsson, CEO and Co-Founder of Modal, explains what drove demand for Modal's sandboxes (vibe coding apps, background agents, and reinforcement learning rollouts) and why he thinks agent harnesses should run outside the sandbox, with sandboxes that are short-lived, maybe one per tool call. He also covers how Modal, which builds AI infrastructure for inference, training, and sandboxes, moved up the stack with auto and shared endpoints for open source models, when he thinks RL is worth its cost over supervised fine-tuning, and the multi-node training and discounted batch products Modal is building. The conversation closes on how Modal's engineers use coding agents and why Erik is watching voice to voice models.
This session was captured at Interrupt New York City, the agent conference by LangChain.
Chapters:
0:00 What Modal builds for inference, training, and sandboxes
0:57 Was Modal built for AI workloads before ChatGPT?
2:47 What made Modal's sandboxes finally take off?
5:35 Why Modal is moving sandboxes from gVisor to VMs
7:11 Should the agent harness run inside the sandbox?
10:09 Do agents need a full VM or a lightweight sandbox?
12:27 Are open source models close enough to swap in?
15:49 When is RL worth it over supervised fine-tuning?
17:08 Why Modal built auto and shared inference endpoints
18:43 Can reinforcement learning be packaged as a product?
19:44 What Modal is building next for training and batch jobs
21:09 Why Modal built a team of forward deployed engineers
23:25 How Modal uses coding agents and MCP servers internally
26:41 Why Erik is bullish on voice to voice models
Resources:
→ Modal: modal.com
→ Modal Sandboxes: modal.com/products/sandboxes
→ Deep Agents with Modal sandboxes: docs.langchain.com/oss/python/integrations/sandboxes/modal
→ Interrupt, the agent conference by LangChain: interrupt.langchain.com










