PS DEV

Session Description

LLM agents are increasingly embedded in real production workflows: marketing automation, customer journeys, ticket triage, personalization engines, and more. But when these agents hallucinate, ignore business rules, or lose track of long-lived state, the consequences are real: revenue leakage, brand risk, legal exposure, and broken user experiences.

This talk introduces Time-Traveling Agents, an architecture that combines LLMs with event sourcing, replay, and deterministic correction. Instead of treating agents as stateful black boxes, we rebuild state exclusively from an event log, capture every decision as an immutable fact, and enable the system to:

• Rewind to any point in history
• Retry decisions with injected corrections or constraints
• Recover safely from invalid or unsafe agent behavior

The result is an AI system that is auditable, rewindable, reproducible, and compliant by design, a requirement for any enterprise deploying LLMs into high-impact domains like pricing, content generation, targeting, or automated decisions.

Attendees will learn:

• Why LLM agents fail in production and how event-driven design fixes it
• How to structure logs, commands, decisions, and corrections for deterministic replay
• How to build systems that safely “time travel” for QA, debugging, and counterfactual analysis
• Practical patterns for integrating LLM agents with Pulsar/Kafka and real-world validators

If your organization is experimenting with LLM agents or struggling with making them safe, explainable, and production-grade, this talk provides a blueprint for building AI systems that can rewind, retry, and recover.

Slide preview: https://danishrehman.com/assets/submission.pdf