Research

Research on how AI agents receive authority, stay within it and answer for every effect.

As AI moves from answering to acting, the hard questions move with it: who authorized this step, what the agent knew when it acted, and how its work is checked, stopped or recovered. We study these questions with formal models, reference implementations and reproducible evaluations, and publish the results as open working papers and an open protocol.

Recent working papers follow agents through changing business rules and evidence: assembling time-consistent context, scheduling around approvals, repairing memory and task graphs, and recovering from partial execution. Others cover tool dependencies, multilingual business semantics, and how authority and budgets carry through delegation.

A water-level gauge standing in a lake beside the same gauge drawn as a scale, its readings charted
24
papers and working papers published
arXivSSRN
Open preprints, free for anyone to read
OpenPort
An open protocol for AI tool access, with its code public. View on GitHub

Latest publications

Recent papers from Accentrust research, presented with their relationship to the systems and projects they extend.

View all publications

Focus areas

Our working papers group around four questions every governed agent has to answer: what it may do, how its effects are checked, what it knows, and how it recovers.

  1. Authority that follows intent

    Authority bound to what the current request justifies, checked again for the resources and subgoals an agent acquires, and revoked cleanly across delegated and asynchronous work.

  2. Effects the server can verify

    Draft-first writes, server-verified action claims and dependency-closure binding, so every tool effect is checked against current facts before it lands.

  3. Context that stays true

    Time-consistent evidence across documents and transactions, context scheduled around approvals, and memory, plans and business meaning repaired as rules and evidence change.

  4. Recovery under real conditions

    Resuming after partial execution and ambiguous tool outcomes, clarifying within a shared question budget, and measuring how memory, approval, execution and recovery interact.

Projects

Projects are where the research becomes working software. OpenPort turns our governance work into an open protocol any runtime can implement.

Fog over a lake beside the same shoreline drawn as a single line with a low sun.

Research philosophy

We study the moment an AI system’s intention becomes an effect, and design the controls that make that moment accountable.

We start with the workflow, not the demo. We study the moments where information is fragmented and responsibility is real—handoffs, approvals, and decisions that have consequences in the world.

We state what each result proves and the conditions it holds under, so others can build on it with confidence.

We treat systems as living. Evaluation, monitoring, and feedback loops are how prototypes become practice—and how they keep improving as data and needs change.

We publish openly: working papers anyone can read, an open protocol with public code, and research with academic partners.

Academic partners

  • The University of British Columbia
  • Georgia Institute of Technology
  • Duke University
  • University of Illinois Urbana-Champaign
  • Simon Fraser University