Autonomous Systems: Building Infrastructure That Thinks and Adapts
Talk 1: Agent Harnesses as Distributed Systems
Long-running agent harnesses are the subject of a growing body of work. A shared set of primitives keeps recurring: deterministic orchestration around probabilistic actors, durable shared state that outlives context windows, and verification gates that catch hallucinated progress before it compounds. Drawing on a year of building an agentic harness that generates and optimizes LLM inference runtimes for Arm hardware from cloud to edge, we walk through how these primitives cash out in practice: an immutable experiment ledger with promotion reasons attached, a knowledge base that must remain consistent across restarts and model upgrades, and correctness gates ordered strictly before performance gates. We describe concrete failure modes — hallucinated correctness, ledger contention, knowledge-base drift — and the distributed systems techniques that kept the harness honest.
Talk 2: Why AI Needs Polystores
Applications evolve. Databases stay frozen in time.
Every application begins with a storage decision. We choose a database based on today’s requirements, then spend years building around that choice. But applications rarely stand still. New features are added, access patterns change, data grows, and entirely new workloads emerge. The database, however, remains constrained by architectural decisions made long ago because changing storage engines is expensive, risky, and often requires rewriting the application.
What if databases could evolve alongside your application? What if they could continuously observe workload patterns, understand application semantics, recommend the best storage engine for each dataset, and transparently migrate data as workloads change?
Until recently, this vision was out of reach. Choosing the right storage engine requires understanding the nuances and tradeoffs of many different storage engines, each with its own architecture, performance characteristics, and operational complexity. Very few engineers possess that breadth of expertise. Recent advances in AI make it possible to capture this knowledge, reason about application workloads, and automate storage selection and migration at a scale that was previously impractical.
In this talk, we’ll explore why polystores are the right architecture for this future. We’ll discuss how AI can analyze workloads, route queries to the most appropriate storage engine, recommend or automate data migrations as applications evolve, and present a single logical database to developers. Finally, we’ll demonstrate these ideas through KalDB, an autonomous polystore that adapts to applications instead of forcing applications to adapt to their databases.
About Speakers
Thomas Cottenier is a Senior Principal Engineer at Arm, working on developer platforms and AI services across edge and cloud environments. He previously served as a Principal Engineer at AWS, where he contributed to AI-powered developer tools including CodeWhisperer, Amazon Q, and Kiro. He holds a PhD in Computer Science in programming languages and software engineering and has over 20 years of experience spanning compilers, code generation and AI acceleration.
Suman Karumuri is a Principal Engineer specializing in distributed systems, databases, and large scale search infrastructure. Over the past two decades, he has built infrastructure at Airbnb, Slack, Twitter, Meta, Amazon, Microsoft, Yahoo, and Goldman Sachs. His work includes leading Airbnb’s next generation observability platform, architecting Slack’s serverless search backend, and co-authoring the OpenTracing specification. He is currently building next generation database infrastructure focused on autonomous polystores that use AI to optimize data placement, query execution, and storage evolution across heterogeneous storage engines.
Thank you Arm for sponsoring the event.