Building Agent Architectures: Design Decisions That Matter

​You can ship an agent in an afternoon. Getting multiple agents to work together in production without losing context, overreaching, or silently breaking—that's a different problem entirely. And most of the decisions that matter aren't about code. They're about scoping.

​Luke Bechtel has built multi-agent systems in production—agentic data pipelines, DAG-based workflows, hybrid RAG context systems—and will break down the design decisions that actually determine whether an agent system works or falls apart.

​What he'll cover:

• ​What an agent actually is at the architecture level, and what it isn't

• ​The two decisions that define every agent system: what can it see, and what can it do

• ​Context window management

• ​Reflexive systems (flexible bot swarms) vs. fixed systems (narrowly scoped repeaters)

• ​Getting multiple agents to coordinate without everything falling apart

• ​Production failure modes and how to avoid them

​About Luke:

​Founding engineer at Ardent AI, where he built agentic data engineering pipelines, DAG-based agent workflows, and hybrid RAG context systems. Currently at Infinity (inference). MS in CS from Georgia Tech, ML specialization.

​Presented by Rally SF.