Your Agents Can’t Be Trusted: Designing for Failures in Agentic Systems

As AI systems become more autonomous, the challenge is no longer simply getting an agent to perform a task. The real challenge is designing systems that remain reliable when agents make the wrong decision, misunderstand context, call the wrong tool, return incomplete information or fail entirely.
Your Agents Can’t Be Trusted: Designing for Failures in Agentic Systems explores the engineering principles behind building resilient agentic applications. The session will look at why failures should be treated as an expected part of agent behaviour, rather than an exceptional case, and how developers can design systems that detect, contain and recover from those failures.
We’ll explore areas such as validation and guardrails, retries and fallbacks, human intervention, observability, tool failures, state management and designing workflows that remain predictable even when the AI itself is not.
If you are building AI agents, integrating LLMs into production systems, or simply interested in how reliable agentic architectures should be designed, this session will provide practical perspectives on building AI systems that can fail safely and recover gracefully.