RESEARCH
Intelligence That Remains Faithful to Intent
Our research explores how neural methods and symbolic reasoning can work together to make AI agents more reliable. We focus on agent fidelity: preserving the purpose, constraints, and standards of the work as agents interpret, decide, and act.
Neural models bring flexibility to language and unfamiliar situations. Symbolic methods provide explicit representations of knowledge, rules, and obligations. We investigate how these strengths can be combined—introducing determinism where decisions must be predictable, while retaining flexibility where judgment is useful.
What Does Faithful Execution Require?
An agent can remain within its permissions and still misunderstand the assignment. It can follow the prescribed steps and produce an inadequate result. Dependable work requires attention to both the boundaries of execution and the purpose those boundaries are meant to serve.
We study how that purpose can be represented, carried through execution, and evaluated against the result. This includes knowing when an agent should continue, when it should seek clarification, and when responsibility should return to a person.
Our Research Areas
Neurosymbolic Reasoning
Flexible reasoning. Explicit foundations.
Language models can interpret requests, recognize patterns, and propose ways forward. Consequential work also depends on facts, relationships, and rules that must remain explicit.
We investigate architectures that connect neural reasoning with structured knowledge and symbolic methods. Our interest is in how these methods participate in decisions: checking a proposed action against a rule, identifying a missing prerequisite, or distinguishing an inference from an established fact.
The research question is where explicit reasoning and constraints can improve predictability—and how to preserve their authority when they interact with probabilistic models.
Specifications for Faithful Execution
Make the intended work examinable.
Instructions often leave important expectations unstated. People supply context, recognize exceptions, and infer what a satisfactory result would mean. Agents need a more deliberate account of those expectations.
We explore ways to represent objectives, permissions, obligations, and acceptance criteria in specifications that people can review and systems can check. These specifications can help connect human intent to the actions an agent proposes and the evidence it produces.
A central challenge is the translation itself. A system may conform precisely to a specification that incompletely captures the intended work. We therefore examine both the specification and execution against it.
Training and Evaluation
Prepare for responsibility. Test the boundaries of readiness.
Our experience training software developers informs how we approach the preparation of agents: establish context, make expectations clear, use examples that expose misunderstanding, and refine instruction through feedback.
We investigate how domain knowledge, worked examples, tool guidance, and evaluation cases shape agent behavior. Evaluation must include ambiguous requests, incomplete evidence, conflicting instructions, and situations where proceeding would exceed the agent’s authority.
We are interested in whether improvements hold across varied conditions—and what evidence should be required before an agent is entrusted with greater responsibility.
Fidelity Across Collaborating Agents
Preserve intent through every handoff.
Work changes hands. A person defines an objective, an agent investigates, another develops a result, and a specialist reviews it. Each handoff creates an opportunity for context to disappear or assumptions to become mistaken for facts.
We study how objectives, constraints, evidence, and unresolved questions can remain available as work moves between participants. This includes how delegated tasks relate to the original assignment and how their results are reconciled.
The aim is to evaluate the contribution of the whole team, including whether locally reasonable actions add up to the work that was actually intended.
Building on Open Research and Specifications
Our research draws on work in neurosymbolic AI, formal methods, agent evaluation, and runtime verification. We examine emerging approaches for their practical contribution to faithful execution.
One relevant foundation is Microsoft’s Agent Control Specification (ACS), introduced by its Responsible AI team. ACS provides a framework for evaluating policies at defined points during agent execution and returning decisions for the surrounding application to enforce.
We investigate how such controls can work alongside representations of purpose, domain knowledge, and acceptance criteria. A policy decision can establish whether an action is permitted under specified conditions. Further evidence is needed to establish whether the overall work was understood and completed correctly.
Research in Practice
Research informs our engineering choices. Engineering exposes the assumptions our research must examine.
We connect these questions to the design of agents, workflows, and review mechanisms for work where quality matters and mistakes have consequences. Two areas of our work make that connection particularly clear.
Agent Engineering & Orchestration
Preparing agents for defined responsibilities means connecting instruction and evaluation with the architecture of execution. Knowledge, tools, permissions, approval points, and recovery paths all influence whether an agent remains faithful to its assignment.
Inspectable AI Reasoning
People need a basis for evaluating an agent’s conclusions. Our ongoing research explores how supporting evidence, assumptions, and explicit rules can be presented in forms that people can examine and challenge.
We distinguish an explanation that sounds plausible from evidence that supports the result.
Bring a Research Question Grounded in Real Work
Where does your work require predictable decisions? What must an agent preserve, demonstrate, or refer to a person?
We welcome conversations with organizations exploring these questions, particularly in regulated and other consequential settings where dependable execution requires more than a convincing response.