KNOW YOUR AGENT / SYSTEMIC AGENT RISK
What is KYA risk?
Know Your Agent starts with understanding an agent: its identity, purpose, accountable owner, authority and operating boundaries. What is it allowed to do, for whom, and under which controls?
From knowing the agent to understanding the system
An individual-agent review can examine the tools and data an agent uses, the actions it can take, the evidence behind its decisions and the person responsible for oversight. That gives a starting point for asking whether its behavior stays within the intended boundary.
At AgentRisk, KYA Risk (KYARisk) is our approach to examining systemic agent risk in financial institutions: how autonomous agents’ authority, decisions and shared dependencies combine across a workflow. We use “systemic” here for exposure across the defined workflow; it does not automatically mean a threat to the entire financial system.
The question is practical: if each component passes its own review, what evidence shows that the combined workflow remains within the institution’s intended exposure?
Individual review and collective exposure
Both views matter. Reviewing the system adds questions about interactions; it does not replace identity, security or individual control review.
| Question | Individual agent | Collective / systemic |
|---|---|---|
| Authority | What can this agent authorize? | What can the agents authorize in combination? |
| Evidence | What supports this agent’s decision? | Do several decisions depend on the same information or each other’s assurances? |
| Controls | Does this agent pass its own limit? | Does a shared control enforce the total boundary before commitments occur? |
| Dependencies | Which services does this agent need? | What happens when several agents rely on the same service? |
| Decision | Can this agent operate as scoped? | What evidence supports the proposed autonomy of the whole workflow? |
Three permitted commitments. One exceeded cap.
ILLUSTRATIVE / ONE COMMITMENT WINDOW
$8M + $8M + $8M = $24M
Each of three agents makes an $8M commitment within its own $10M limit. The institution intends a $20M combined cap for the same window. The combined $24M therefore exceeds that boundary by $4M.
The $4M excess is a commitment-boundary breach, not a measured loss. These are hypothetical figures, not a customer result or a prediction.
A review would ask whether the cap is defined consistently, whether all relevant commitments enter the shared record, and whether the control can reserve capacity before another agent commits. Concurrency, delayed updates and exception paths are questions to examine in the actual workflow, not failures established by this illustration.
How we examine a workflow
AgentRisk’s Systemic Exposure Review is scoped around one consequential workflow and one autonomy decision. Our review method is map, challenge, test, decide.
Map
Define the workflow, agents, authority, owners, dependencies and intended exposure boundary. Record the evidence and what remains unknown.
Challenge
Examine plausible paths through shared information, combined commitments and common services. State the assumptions and conditions behind each scenario.
Test
Where an appropriate test environment and evidence are available, check whether controls interrupt the scoped paths before the boundary is exceeded. Distinguish observed behavior from modeled or hypothesized outcomes.
Decide
Set out evidence, limitations and options: approve within scope, restrict authority, change controls, request more evidence or defer. The institution owns approval and risk acceptance.
Changes in permissions, models, data, controls or dependencies may change the conditions behind an earlier decision and warrant reassessment.
Relevant references and their limits
NIST’s AI Risk Management Framework 1.0 Core addresses system context, third-party dependencies and testing. It provides general AI risk-management guidance; it does not define or endorse AgentRisk’s offering.
Guy Burstein’s geometric vector framework for high-dimensional interaction modeling is the research reference linked from our homepage. It is research background, not evidence that deployed autonomous-agent assessments have been validated.
For related operational-risk research, see Burstein and Zuckerman, “Uncertainty Reduction in Operational Risk Management Process” (2024). Research findings do not establish the effectiveness of a control in your workflow; that requires appropriate evidence and testing.
Frequently asked questions
What does KYA mean here?
KYA means Know Your Agent: understanding an agent’s identity, purpose, accountable owner, authority and operating boundaries. At AgentRisk, KYA Risk extends that inquiry to the workflow and its combined exposure.
How is systemic agent risk different from individual-agent risk?
Individual review asks whether one agent operates within its authority and controls. Systemic review also asks what agents can collectively cause through combined commitments, shared information or common services.
Is the $4M in the example a loss?
No. Three $8M commitments total $24M against an intended $20M cap. The $4M is an illustrative commitment-boundary breach, not a measured or realized loss.
Does KYA Risk certify that a workflow is safe?
No. The review is intended to provide scoped evidence, control findings and decision options. The institution retains responsibility for approval and risk acceptance; results depend on the evidence and conditions examined.
SYSTEMIC EXPOSURE REVIEW
What is changing in your workflow?
Bring us a proposed autonomy change and the boundary that must hold. We can discuss the scope and evidence needed for a review.
Discuss a Systemic Exposure Review