Patrick Dougherty is the founder of Aster Labs and co-founder and former chief technology officer of Rasgo, an enterprise analytics company acquired in 2024. He builds AI agents that can investigate enterprise data, recognize when information is insufficient, and operate within the permissions of the people using them.
From enterprise analytics to autonomous agents
After earning a master’s degree in analytics, Dougherty worked as a data scientist at Dell, developing models to improve sales and marketing returns. At Slalom, he led its Atlanta data science practice and subsequently its Charlotte data practice, confronting the gap between companies’ AI ambitions and the condition of their underlying data.
In 2020, he co-founded Rasgo with Jared Parker, leading product and engineering while Parker served as chief executive. Rasgo initially helped data scientists prepare warehouse data for machine learning; Dougherty emphasized rapid experimentation, customer feedback, and products suited to organizations’ actual capabilities. His case against centralized data science extended that philosophy: embedding data scientists within business functions keeps their work accountable to concrete decisions and results.
As language models improved, Dougherty rebuilt Rasgo around agents capable of exploring enterprise data, querying warehouses, and revising their approach. Following the company’s 2024 acquisition, he founded Aster Labs and announced its launch with Andrew Griffin, identifying applications including insurance-claim appeals, investment documents, software support, and invoice classification.
Principles for agents that actually work
Iterative warehouse exploration: Dougherty defines an agent by its ability to pursue an objective, use tools, and independently choose its next action. Instead of loading every available table into a prompt, he equips agents to search schemas, inspect columns, and gather relevant information as questions develop.
Recognizing unanswerable questions: Given Salesforce tables without a reliable customer-churn indicator, GPT-4o generated plausible but unsupported SQL, while OpenAI o1 recognized that the question could not be answered from the available data. Dougherty treats that restraint as essential to trustworthy analytics.
Agent-computer interface: Tool definitions and response formats directly influence reliability: JSON improved GPT-4o’s interpretation of warehouse results, while XML worked better for Claude. Repeated malformed calls can signal that a tool’s design conflicts with a model’s expectations. His production-agent lessons also caution that fine-tuning an agent’s primary reasoning model can make its behavior overly rigid.
User-scoped enterprise permissions: Agents should inherit each user’s actual warehouse permissions through mechanisms such as OAuth. Dougherty locates durable product value in secure integrations, usable interfaces, and operational control, and cautions against frameworks that obscure authentication or complicate production debugging.
Manager-worker agent hierarchies: A manager should own the final outcome while delegating bounded work to specialists. Dougherty generally favors teams of roughly five to eight agents; larger groups increase coordination failures, loops, and irrelevant work.
