Dynamic runtime governance of AI agents through supervisory agents - ON-1255

Project type: Research
Desired discipline(s): Engineering - computer / electrical, Engineering, Computer science, Mathematical Sciences, Operations research
Company: FairFuture AI Inc.
Project Length: 6 months to 1 year
Preferred start date: 01/04/2027
Language requirement: English
Location(s): Toronto, ON, Canada
No. of positions: 1
Desired education level: Master'sPhDPostdoctoral fellowRecent graduate
Open to applicants registered at an institution outside of Canada: No

About the company: 

FairFuture AI is a Canadian company focused on Responsible AI research and development and building an enterprise AI governance platform. The platform is designed to help organizations maintain an AI inventory, assess risk, manage controls and evidence, and adapt oversight as models, data, policies and AI agent behaviour change. Fiara is the intelligence layer intended to simplify this governance experience.

Our ambition is to develop a global Responsible AI platform and a Canadian research hub through sustained academic and industry collaboration. These three projects address immediate commercialization and technical research priorities within that longer-term program. We aim to create commercially useful intellectual property in Canada, develop skilled Canadian talent, support future employment and build export opportunities. Research outputs are intended for incorporation into FairFuture's platform, subject to agreed academic IP, licensing and publication arrangements.

Describe the project.: 

FairFuture will have a foundational platform and basic agent orchestration before this internship. This Accelerate project will investigate how Fiara's supervisory AI agents can assess and govern other agents during execution as tasks, permissions, models and policies change. The research question is whether a hybrid architecture combining deterministic enforcement with evidence-grounded supervisory agents improves risk detection and intervention over static controls without unacceptable latency or unnecessary interruption.

The intern will investigate delegated authority, cross-agent handoffs, policy conflicts, uncertain evidence and changes during a running workflow. Two representative enterprise workflows will bound the experiments. Candidate designs will separate observation, policy interpretation, authorization and intervention; version policies and decisions; and restrict the supervisor's own authority. Hard permission boundaries will remain deterministic, with human approval for designated consequential actions.

Methods will include a focused literature review, threat modelling, executable policy representations and controlled fault injection. Experiments will compare static rules, an LLM-only supervisor and the hybrid design across two agent frameworks. Test cases will cover prompt injection, unauthorized tools, excessive delegation, policy updates, poisoned observations and supervisor failure. Metrics will include missed violations, false interventions, task completion, decision traceability, recovery, latency and inference cost, with repeated trials and ablation studies.

Months 1–2: research gaps, hypotheses, policy/event contracts and benchmarks. Months 3–4: alternative prototypes. Months 5–6: comparative evaluation and refinement. Months 7–8: bounded integration into FairFuture and reproducible validation.

Deliverables will include reusable supervisory components, deterministic enforcement interfaces, a benchmark harness, evidence-linked decision records and documented integration contracts. The research targets transferable methods beyond routine orchestration. Weekly company and academic supervision will support experimental design and implementation. The project aims to create Canadian IP and research skills, with platform-use rights agreed with the academic institution before work starts.

Required expertise/skills: 

Required: strong Python; agentic AI and LLM evaluation; distributed systems and API design; experimental design, statistical analysis and reproducible benchmarking; secure software engineering; ability to implement and compare research prototypes. Candidates should understand tool permissions, prompt injection, stateful workflows and the limits of probabilistic model outputs.

Preferred: a research master's or PhD student in computer science or computer engineering with work in multi-agent systems, AI safety, runtime verification or security. Assets include policy engines, formal methods, constrained planning, uncertainty estimation, asynchronous orchestration and observability. Experience with containers, PostgreSQL and version-controlled testing is useful. Components should support provider-independent deployment. The academic supervisor should contribute expertise in agent systems, verification or AI security; FairFuture's technical lead will mentor platform integration. Basic UI development or ordinary agent wiring alone is insufficient for this research scope.