Research overview · Kiran Yenaganti

Evidence-Bound Runtime Governance for Autonomous Multi-Agent AI Systems

Policy Enforcement, Monitoring, and Risk-Based Human Escalation

An AI agent can have permission to act and still lack the evidence to make a sound decision.

Research manuscript · Read the full text on ResearchGate. Publication and validation details are summarized below.

The problem

Permission does not settle the question.

Consider an AI leasing assistant asked to schedule a tour. It may be authorized to use the scheduling system, but the availability it retrieved could be out of date. Another agent may have supplied an assumption rather than a confirmed fact.

The policy check answers whether the action is allowed. It does not, by itself, establish whether the information supporting that action is sufficient.

This is the distinction behind my work on evidence-bound runtime governance: evaluating the basis for an action while a workflow is running, alongside its permissions and approval requirements.

Three questions at execution time.

Is the action permitted?
Check the proposed action against the applicable policy and the agent’s authority.
What supports the decision?
Examine whether the information is relevant, current and sufficient, and whether important gaps or contradictions remain.
Does a person need to decide?
When the evidence or authority is insufficient, pause or escalate rather than treating uncertainty as approval.

From a use case to a research question.

The virtual leasing concept led me to broader questions about how specialized agents coordinate, use enterprise data and recover when a model or information source fails. Those questions also informed my patent application on governed multi-agent workflows.

HOMININI provides a separate engineering case study: a diagnostic repair that explains why execution stopped without silently restarting work. Its developer-run regression tests cover that specific diagnostic behavior.

Explore the HOMININI case study →

Scope and status.

This is a research manuscript shared on ResearchGate. Journal acceptance and peer-review status have not been verified for this website. HOMININI’s diagnostic tests are a separate engineering result, rather than validation of the full proposed approach. Broader performance and effectiveness require empirical evaluation.

Useful questions include how evidence quality is defined, how escalation thresholds are chosen, what happens when sources disagree, and how the governance checks affect task completion and execution cost.

Source manuscript