Most agentic systems are built and evaluated for single sessions. Institutional finance workflows do not work that way, they run for days or weeks, span multiple agents and tool calls, and need to recover from failure without a human stepping in every time something drifts. This session covers what it takes to design agentic systems for long-running financial workflows, from state and memory architectures that survive across sessions to self-healing patterns that detect drift, retry intelligently, and escalate only when they genuinely need to. It looks at where these systems tend to fail in production and what separates a resilient long horizon agent from one that quietly compounds errors over time. Key takeaways: - How to design memory and state layers that hold up across long-running, multi-step financial workflows - Self-healing patterns for detecting and recovering from agent drift and tool-call failures without constant human intervention - Where long horizon agentic systems break down in production and how to design around it - How to decide what should trigger automatic recovery versus escalation to a human