Making LLM Agents Think Once, Then Act Fast

PowerTalk

About the session

LLM agents often fail at enterprise scale because we treat them as data processors rather than orchestrators. Facing petabytes of production telemetry and billion-row tables, the standard "Step-and-Loop" architecture collapses under latency, security risks, and context window bloat. This talk explores a shift from reactive agents to declarative planners. We detail how we built a high-scale system that separates reasoning from execution. By using jq as a safe, sandboxed transformation layer, our agents generate complete, auditable execution plans upfront keeping massive datasets out of the context window while maintaining strict security.

Speaker

Download Brochure