Agentic systems fail in ways regular request response systems often do not. A workflow can be interrupted halfway through, a tool call can time out, a provider can hit capacity, a retry can duplicate work, or a human may need to approve a step before execution continues.

In this session, I’ll walk through the reliability patterns that make agentic systems production ready. We’ll cover:

• How workflow state makes every agent request inspectable
• How event logs show each agent step, tool call, output, and failure
• How checkpoints let agents resume completed work after interruption
• How replay helps recover workflows without starting from the beginning
• How leases and recovery streams allow another worker to safely continue unfinished work
• How circuit breakers prevent repeated calls to failing providers
• How distributed semaphores protect shared provider capacity
• How dead letter queues keep failed steps visible after retries are exhausted
• How human approvals pause and resume workflows around sensitive actions

We’ll use a stock analysis agent as the running example and examine the failure modes that appear when agents call tools, depend on external APIs, coordinate across workers, and run longer than a single request.

Redis will appear as one practical implementation layer for fast shared state, streams, locks, queues, and recovery metadata. The main focus is the reliability model: how to preserve state, expose failures, control retries, recover safely, and keep agentic workflows understandable in production.

Participants will leave with a practical mental model for building agentic systems that can be observed, resumed, throttled, recovered, and operated under real production conditions.

This session covers reliability patterns for agentic systems: workflow state, event logs, checkpoints, replay, recovery, circuit breakers, semaphores, dead letter queues, and human approvals. We’ll use a stock analysis agent to show how agents can preserve state, recover from failures, control tool usage, and stay inspectable in production.