Architecting multi-agent systems

Observability, typed contracts and enterprise scale

The shift from single-agent prototypes to multi-agent production often represent a distributed systems challenge first and an AI challenge second. How can teams build reliable multi-agent systems with the right control and oversight?

This XConf 2026 playlist brings together four talks that looks at the PoC-to-production cliff. They show how teams can move beyond prompt chaining and toy ReAct loops to address the architectural discipline required for production: runtime execution traces, typed contracts with confidence-score gating to stop error cascades, and scaling multi-agent deep research on high-performance infrastructure.

Watch the playlist below to ship multi-agent systems with confidence.

Be the first to hear about XConf 2027

Inside this playlist

But it worked in my PoC!

What does it take to automate enterprise IT at scale with AI agents? More than expected. This talk shares the hard-won lessons from deploying agents to manage software updates across thousands of devices globally. Discover what works and learn why multi-agent systems are a distributed systems problem first and an AI problem second.

What agents do when you're not looking

Coding agents have evolved fast. This session traces the evolution, shares lessons learned from building and observing agent teams in Claude, and closes with the emerging practitioner playbook. Learn how teams can observe, evaluate and govern agents in order to use them with more confidence in real-world delivery.

Agent don't fail, they guess

Multi-agent pipelines fail differently from single agents. One agent gets it wrong and you see it. A pipeline compounds wrong assumptions invisibly. This session shares the experience of building and deploying two agents: One analysing, one transforming. Each has to pass a typed contract between them and a gating mechanism stops the pipeline when certainty falls below threshold.

Deep research at scale

Scaling LLMs to 15M+ domain-heavy documents exposes the limits of standard RAG. Complex enterprise queries demand multi-hop reasoning, multi-document synthesis, and strict domain accuracy. The session shares the blueprint for taking a multi-agent deep research engine from prototype to production, and lessons from MLflow observability, streaming logs, and Bedrock infrastructure.

Why you should watch

All-knowing orchestrators or smaller agents?

Get lessons learned from two failed attmpts in automating enterprise IT with AI agents. Find out why monolithic orchestrators fail under 3M tokens and how dynamic, specialized sub-agents fix them.

Runtime observability for agents

Diagnose unexpected behaviors, infinite loops, and delegation failures through execution traces rather than static demos to build agent teams with confidence.