Michael Albada is a platform architect at NVIDIA developing agentic systems for chip design. Previously a principal applied scientist working on Microsoft Security Copilot, he wrote Building Applications with AI Agents, an O’Reilly guide to making autonomous applications reliable enough for production.
Earlier in his career, Albada worked on natural-language processing at ServiceNow, built machine-learning systems for geospatial intelligence at Uber, and worked in startups. At Microsoft, he developed cybersecurity agents for planning, orchestration, and skill selection. His research on ExCyTIn-Bench addresses a difficult practical question: whether agents can investigate threats by navigating heterogeneous security logs and following multistep chains of evidence. He subsequently joined NVIDIA, bringing his experience with complex, consequential agent workflows to semiconductor engineering.
Albada holds degrees from Stanford University, the University of Cambridge, and Georgia Tech. His 2025 book covers model selection, tools, memory, orchestration, evaluation, reliability, human oversight, and multiagent coordination.
Principles for dependable agents
Effectiveness before autonomy. Agentic behavior is useful only when it improves results. Albada favors simple chains and branching workflows when they are cheaper, more reliable, and easier to measure, adding autonomy when fixed workflows become too brittle or difficult to maintain.
Small, coherent tool surfaces. Exposing every available API to a model increases confusion and tool-selection errors. He recommends narrowly scoped, clearly described actions; when capabilities multiply, specialized agents can manage related tools under a coordinating system.
Deterministic safeguards around probabilistic models. Application state and consequential business rules belong outside the model. Conventional software should validate intermediate updates and enforce explicit conditions before sensitive actions proceed.
Evaluation-driven agent development. Human-reviewed test cases, carefully constructed synthetic examples, and systematic failure analysis make decisions about models, memory, tools, and orchestration measurable. Cybersecurity benchmarks test whether agents can actually assemble investigative evidence, not merely generate convincing answers.
Albada also treats tracing, adversarial testing, and human escalation as essential production infrastructure. His AI Engineer World’s Fair talk highlights practical approaches including OpenTelemetry, OpenLLMetry, DSPy, and Microsoft’s PyRIT.
A useful agent needs more than tool access: clear interfaces, maintainable orchestration, deterministic safeguards, and an evaluation loop that turns failures into improvements.
Why does a promising prototype remain hard to ship?
