Reliability patterns for agentic systems

Production systems also need reliability patterns that let agents recover from failed tool calls, unavailable services, rate limits, stale state, duplicate work, and partial execution.

This session covers reliability patterns for agentic systems using Redis 8.

We will look at durable state, Redis Streams for event history, queues for async work, TTLs for bounded memory, locks for coordinated actions, rate limiting for model and tool calls, dead letter queues for failed steps, checkpoints for recovery, and replay for failed workflows.

• Streams: Record agent events, tool calls, tool results, retries, and state changes as an ordered history.
• Queues: Move slow or unreliable work out of the main agent path and process it asynchronously.
• Retries: Retry failed model calls, tool calls, and jobs with limits, backoff, and stop conditions.
• Dead letter queues: Capture work that keeps failing so it does not block the rest of the system.