Agentic AI
Build Reliable AI Systems with Constrained Tools, Not Open-Ended Agents
A former startup founder turned NVIDIA engineer shares lessons from building production AI systems: reliability comes from splitting tasks, constraining what AI can do, and combining deterministic tools with agentic discovery.
InfoQ
The Allure (and Risk) of Adding AI to Everything#
Businesses face constant temptation to inject large language models into existing products and processes without careful thought. The idea sounds appealing—throw more tokens at a problem and discover new possibilities. But applying AI indiscriminately often leads to unreliable, erratic systems that produce worse outcomes than the systems they replaced.
Aaron Erickson, now at NVIDIA, learned this lesson firsthand. Years earlier, while building an HR tech startup around organizational charts, his team faced funding pressure in 2023 to "add AI." The result: they built a ChatGPT plugin that could help restructure organizations. It actually worked—the system could analyze company hierarchies, propose flattening changes, and generate reorg emails. But that didn't matter. The business failed anyway. "We didn't become the future of HR," Erickson notes with characteristic candor.
From Org Charts to GPU Fleets#
The principles that failed in HR turned out to be transferable. At NVIDIA, Erickson's first project wasn't AI-driven at all—it was building a resource allocation system for GPU clusters, where thousands of research teams compete for scarce hardware. The domain shift revealed striking parallels: organizations have employees and hierarchies; GPU systems have jobs and infrastructure regions. Both required constraint-solving under resource scarcity.
This realization prompted Erickson's team to build LLo11yPop (a deliberately awkward acronym for "large language model for observability"), an agent architecture designed to troubleshoot and optimize massive GPU clusters.
The Architecture: Specialization Over Generalization#
Rather than giving the LLM free rein to reason about GPU health from scratch, the team applied strict division of labor:
- Retrieval agents know exactly one query type—convert a question into SQL, Elasticsearch queries, or API calls. They're given a few good examples and constrained to a single table or endpoint.
- Analyst agents don't touch databases; they focus on domain-specific reasoning (e.g., "how do H100s fail?"). They reason through problems using retrieval agents as data helpers.
- Orchestrator agents coordinate analyst agents toward a concrete goal, surfacing recommendations to relevant stakeholders.
