Know Your Principal.
The generalization of Know Your Agent to every kind of autonomous actor: AI agents, service accounts, drones, robots, PLCs, and cyber-physical systems.
What it is
Know Your Principal, or KYP, is the general form of the Know Your Agent framework. Where KYA describes the primitives for governing AI agents at runtime, KYP applies the same primitives to any principal that dispatches an action, whether it is a human, an LLM-based agent, a service account, a drone, a robot, a PLC, a controller, a sensor, an actuator, or an autonomous system.
The definition of a principal is deliberately wide. If it can send a request, take an action, or move a piece of state, it is a principal, and if it is a principal, KYP says: give it a signed identity, bound its authority per action, and write a tamper-evident record of what it did.
Why the generalization matters
The near-term pressure on runtime governance comes from AI agents, but the medium-term reality is broader. Enterprise, defense, and industrial systems are increasingly multi-principal by design. A workflow already routinely spans a human operator, one or more agents, a bank of service accounts, a fleet of sensors, and often a physical actuator on the other end.
If governance stops at "AI agents" it leaves the cyber-physical half of the workflow ungoverned. A drone dispatched by an agent, a PLC updated by a workflow, a robot arm actuated on a scheduler, each of these is a principal, and each of them can be the endpoint of an attack chain, a delegation error, or a policy violation. KYP is the framework that treats them as first-class principals, not out-of-scope endpoints.
The formulation matters most in defense, industrial, and critical-infrastructure settings, where the population of principals is not just AI-shaped and where accountability requirements are strictest.
The principal kinds KYP covers
The KYA paper credits the generalization by naming principal kinds beyond just LLM agents. A partial list:
- Human operators, with delegated scopes for high-authority actions
- LLM agents, both single and multi-agent chains
- Service accounts and workload identities
- Autonomous drones and unmanned systems
- Ground robots and manipulation arms
- PLCs and industrial controllers on OT networks
- SCADA supervisory nodes and RTUs
- Sensor networks, dispatchers, and edge devices
- Actuators, valves, and physical effectors
- Autonomous vehicles and mobility platforms
- Ground stations and command relays
- Model runners and inference gateways
- Workflow schedulers and orchestrators
- Data agents and ETL workers
- Cyber-physical fusion nodes that mix compute and control
The exact taxonomy is a moving target. The point is that KYP does not privilege the LLM case. Every entry above is a principal, and every one of them gets the same treatment: signed identity, bounded authority, evidence.
How KYP shows up in Veldt
Two pillars of the Veldt platform carry most of the KYP weight.
- Identity. Portable, verifiable non-human identity that does not assume the principal is a language model. The same identity primitives cover a service account, a drone, and a PLC.
- Authority. Per-action authorization and delegation with only-tighten composition, applied across any principal type.
The runtime governance pipeline is agnostic about which principal kind produced the action. It scores identity, authority, policy compliance, and behavior on every request, and returns one of five verdicts.
Multi-principal delegation
In a multi-principal world, delegation chains rarely stay within one principal kind. An agent hands off to a service account, which triggers a scheduler, which dispatches a drone. If the chain is to remain accountable, delegation attribution has to trace through every hop, regardless of principal type.
KYP is what makes that trace legible. Each hop, agent to service account, service account to scheduler, scheduler to drone, is a signed delegation event with a bounded scope. The chain composes under only-tighten, so no downstream principal can widen authority. The evidence trail records the whole path.
The result: an autonomous drone action is provable back to the human whose authority started it. The same math that works for agent-to-agent delegation works for agent-to-machine and machine-to-machine.