Know Your Agent.

The framework and research foundation behind Veldt: signed identity, continuous trust scoring, delegation attribution, and tamper-evident evidence for every autonomous agent action.

What it is

Know Your Agent, or KYA, is a research framework for governing autonomous AI agents at runtime. It answers a question that traditional identity and access management does not: who authorized this specific action, on this specific principal, under this specific scope, right now?

The framework, formalized in the KYA paper, describes five architectural primitives that, taken together, make an autonomous action provable. Every agent action carries a signed identity, a bounded scope, a delegation chain back to a human principal, a trust verdict, and a tamper-evident evidence record. Nothing else counts as governed.

Why "Know Your Agent"

The name is a deliberate echo of Know Your Customer (KYC) in financial services. KYC exists because banks cannot execute anonymous transactions at scale. The regulatory answer was a framework: verified identity, ongoing due diligence, and a record trail.

Autonomous AI agents sit in the same position banks were in forty years ago. The industry has shifted from answering-AI, models that produce text on request, to acting-AI, agents that write to systems, move money, dispatch machines, and read regulated data. An acting agent without a signed identity, a bounded authority, and an evidence record is the AI equivalent of an anonymous transaction.

KYA is the framework that closes the gap. It gives every agent a verifiable identity, every action a signed scope, and every incident a traceable chain back to a human decision.

The KYA primitives

The paper introduces five architectural primitives. Each one is a load-bearing part of the framework.

• Signed identity. Every agent has a portable, verifiable non-human identity, not a shared API key or ambient service credential.
• Per-action authorization. Authority is bounded, scoped, and checked per action. There is no ambient permission to be an agent.
• Continuous trust scoring. A live behavioral score per principal drives one of five verdicts on every action, not a static allow/deny list.
• Delegation attribution. Every action traces back, through only-tighten composition, to the human principal who started the chain. See delegation attribution.
• Tamper-evident evidence. Every verdict is written to a cryptographically chained record that any auditor can verify offline.

None of the five stands alone. The framework's guarantees hold because all five are enforced on the same action.

How KYA relates to KYP

KYA was formalized around AI agents because that is where the industry pressure is loudest. But the same primitives apply to any autonomous principal, not just LLM-based agents.

Know Your Principal (KYP) is the generalization: the same signed identity, bounded authority, and evidence trail applied to service accounts, drones, robots, PLCs, sensors, and other autonomous systems. When the paper talks about agents, the machinery is agent-shaped. When you generalize it to every dispatching principal, you get KYP.

Put simply: KYA is the framework in its LLM-agent form. KYP is what you get when you let go of the assumption that the principal is a language model.

How Veldt implements KYA

Veldt is the runtime that carries the framework into production. Each of the five primitives maps to a platform pillar:

• Identity. Portable, verifiable identity for every principal.
• Authority. Per-action authorization, delegation, and only-tighten composition.
• Runtime Governance. Continuous trust scoring and the five-verdict pipeline (allow, redact, throttle, escalate, block) in under 50ms.
• Evidence. Signed, tamper-evident audit trail, mapped to 32 regulator packs.
• Attack Chains. Multi-step attack detection over the signed action stream.

KYA remains a framework and a research foundation. Veldt is the implementation. The paper is the source of the definitions on this page.