What infrastructure matters most for long-running AI agent tasks?

AM

Andrea Malagodi

Chief Information Officer · Sonar

Context retrieval

Andrea Malagodi says context retrieval is important, pointing to Vortex, which lets an agent pull context from a codebase—coding guidelines or its graph—to reduce the LLM's wasteful rereading of code.

The answer starts at 13:35.

Rama Akkiraju

VP, Enterprise AI · Nvidia

Observability

Rama Akkiraju says observability is key, and says understanding agent behavior before certifying an agent is critical.

RL

Rodrigo Liang

Co-Founder & CEO · SambaNova Systems

Sustainability

Rodrigo Liang says sustainability is one of the biggest infrastructure challenges, so businesses aren't in a bubble and can build and invest for the long term.

Katelyn Lesse

Head of API Engineering · Anthropic

Durable harness

Katelyn Lesse argues for a durable harness—running the agent's brain on a durable server—so that if an ephemeral sandbox dies, the whole agent doesn't.

Ryan Lopopolo

Member of Technical Staff · OpenAI

Robust agent harness

Ryan Lopopolo says the agent harness becomes increasingly important, calling it the environment humans build to give agents the tools, context, guardrails, and trust needed to translate existing model capability into reliable real-world action.

Aditya Khandelwal

Member of Technical Staff · Amazon

Self-healing capability

Aditya Khandelwal says you need a self-healing system because slop is inevitable, requiring a pipeline to detect it, remove it, and close the loop.

JS

Justin Smith

Founding Product Engineer · Resolve Ai

Continuous learning system

Justin Smith says an underlying learning system is what a model needs to succeed at its task, because it captures understanding of how your system operates and grows as your system evolves.