Self-Improving Agents with test-time compute for FDA

Current artificial intelligence (AI) agents deployed in enterprise settings make static, one-shot decisions and cannot adapt as context changes or learn from the outcomes of past actions. This research addresses two limitations of enterprise AI agents: (i) the inability to dynamically adjust reasoning effort based on the uncertainty and stakes, and (ii) the inability to improve their decision-making capabilities over time by learning from observations. Both result in agents that are brittle under distributional shift, poorly calibrated in their confidence, and unable to close the loop between prediction and consequence.
ServiceNow aims to build Foresight-Driven Agents (FDA): enterprise AI agents that maintain a dynamic world model of an organization, reason over evolving context, and forecast the consequences of decisions. The specific challenges this project addresses are: (a) how to allocate additional test-time reasoning to highuncertainty, high-stakes decisions, and (b) how to design self-improvement mechanisms that allow the agent to update its world model, recalibrate its confidence, and refine its decision strategies based on delayed and noisy outcome signals.
ServiceNow develops a platform to manage and automate large-scale enterprise processes across diverse industries. In 2020, ServiceNow acquired ElementAI to boost its standing in AI research and the Canadian AI scene, leading to enhanced AI-powered products and platform features. ServiceNow AI Research has made significant contributions to the field of foundation models and has a strong presence in developing AI agents for different domains.
This project will advance ServiceNow’s capacity to deploy agents that more reliable, well-calibrated decisions in dynamic environments. By enabling agents to allocate reasoning effort where it matters most and to improve from experience, the project directly addresses operational efficiency. The resulting research contributions, including open-source benchmarks and publications at top venues, will strengthen the Canadian AI research ecosystem and ServiceNow’s position as a leader in enterprise AI.

Faculty Supervisor:

Jose Dolz

Student:

Partner:

ServiceNow Canada

Discipline:

Computer science

Sector:

Professional, scientific and technical services

University:

École de technologie supérieure

Program: