KTP/
github
created 14 November 2026 · last modified 2026-11-14
the four questions · an evidence index

KTP and agent identity.

The agent-identity field sorts projects by four questions: which agent is it, who runs it, who does it act for, what is it allowed to do. The form is credential-shaped. Each question expects an answer that is issued once and carried.

KTP answers all four and half-fits every cell, because the form has no column for the two things the protocol is built on: time (the moment of the grant against the moment of the act) and exposure (what the party behind an agent stands to lose). This page answers the four questions in KTP's terms, says where each answer parts from the form's assumption, and links every claim to a tag-pinned document or a public decision record. Nothing on this page is a conformance claim.

Pinned to the v2.1.0 tag of nmcitra/ktp-rfc. A row marked “on main” is text no release yet carries; a row marked “proposed” is an open issue on the repository’s public clock.

the form’s four questions, answered
01

Which agent is it?

An anchored trajectory. KTP names an agent with an agent URI, but it does not identify the agent by the name. The identity is the Trajectory Chain: an append-only, cryptographically linked record of what the agent has done, co-signed by the zone's Trust Oracle, with one anchored head. A key proves possession. The chain proves continuity.

where it parts from the form

The form expects a name or a key that resolves to an entity. KTP calls that the Passport Fallacy: possession of a credential is a claim about the past, not an identity. Re-identification is preserved. The credential is demoted to a handle.

02

Who runs it?

KTP answers: Who answers for it?

The External Root: a party outside the agent-trust graph that is, in the specification's words, “externally accountable, exposed to loss, and able to revoke.” Every agent's standing includes a term from that root. The chain behind it must end, within a declared hop bound, at “a physical-presence attestation, a legal entity, or a named human.” An agent with no current root computes that term as zero.

where it parts from the form

The form asks who operates the agent. KTP asks who stands to lose, and the two can differ: the proposed repair document says the party who owes for an ended actor is “the root that held authority, never an operator who directed without the power to stop.” An open proposal makes the root a person, with an employer, a licensing body or a board behind them.

03

Who does it act for?

KTP answers: Who vouches for it, and who is served?

Two roles, not one. A sponsor stakes its own standing on a new agent (“Liability, not recommendation”) and keeps a share of the responsibility after the bond closes. A managed agent's principal is the party it acts for, and may contest a ruling afterward. Delegation narrows authority; it “never transfers qualification, capacity, person type, or readiness.”

where it parts from the form

The form has one principal: the party who granted the token. KTP splits the one who vouches and pays from the one who is served and may contest. A stop inside a managed relation is proposed to carry its destination, so a denial never leaves the principal with nobody to receive it.

04

What is it allowed to do?

KTP answers: What can this environment carry, right now?

Whatever the environment can carry at the moment of the act, and no more. Before every action the zone evaluates A ≤ E: the action's risk against the current Trust Score, E_trust = E_base × (1 − R), where R is measured from the environment. A Trust Proof lives at most ten seconds. Trust Tiers permit classes of action; the only numeric bound is the law itself. When A > E the action does not run, and no approval relaxes a correctly evaluated veto. Explanation, review and correction remain, and a correction leads to a fresh evaluation.

where it parts from the form

The form expects a scope that holds between asks. KTP calls that the Static Fallacy. Its answer is a verdict for one act under measured conditions, recomputed; this cell cannot be filled with a permission list.

layer · stage · portability · building blocks

Layer

Authority and delegation, in the form's categories. What sets it apart from other authority layers is where the authority comes from: it is derived from the environment and recomputed, “instead of being settled once at grant time.”

Two cautions. KTP carries a standing beneath the authority (Base Trust, earned through Proof of Resilience) that looks like reputation but is one zone's measure: a measure never crosses a federation link, and a score presented outside its zone has no defined meaning. And KTP has a naming scheme (agent URIs) that is not the identity.

Stage

Draft specification. v2.1.0 “Gödel”, 3 September 2026, archived at DOI 10.5281/zenodo.22293630: 27 documents, five formatted as Internet-Drafts, seventeen JSON Schemas, a public decision log with clocks, and conformance vectors in the repository. v3.0.0 “Ostrom” is planned for 10 December 2026 on a public release proposal.

Three public implementations, each experimental by its own maintainer's description and pinned to v2.1.0. None is a reference implementation. None has run a published vector suite against an authorizer in public. In the maintainer's words on the release proposal, “no claim above ‘specified’ or ‘proposed’ exists for any of it today.”

Portability

An open specification under Apache 2.0. There is no vendor network to be inside of. Standing earned in one zone is zone-local by design; what an agent carries out of a zone is an Exit Attestation, and what a receiving zone makes of it is the receiving zone's measure.

Trust that travels between zones is the subject of KTP-FEDERATION, which at v2.1.0 is largely descriptive. The binding rule (a federation moves decisions about the federation, never a zone's evaluation of an act) and what may cross a link are proposed for v3.0.0. The honest phrase is “specified in draft,” not “federated.”

Building blocks

Specified as bridges, each with a declared ceiling on the standing it can map to and a rule of no trust amplification: OAuth 2.0, OIDC, SAML 2.0, mTLS and API keys (KTP-LEGACY); SPIFFE/SPIRE identities (KTP-MIGRATION). The Trust Proof is a JWT (RFC 7519); canonical JSON is JCS (RFC 8785).

Not specified: MCP, A2A, x402, Web Bot Auth. An MCP binding profile is deferred until after v3.0.0. How KTP reads MCP and A2A as authority models is on Prior Art.

what an outside reader can check

A project sorting itself into a form establishes nothing. These are the artifacts the claims above resolve to.

  • The v2.1.0 tag · every link on this page marked shipped resolves inside it
  • The decision log · every change on a published clock; nothing passes by default
  • Conformance vectors · deterministic cases; none certifies a runtime
  • Implementations · three, experimental, pinned to v2.1.0; the status lines are the maintainers' own
  • Prior Art · the comparison matrix; KTP scored by the same rubric as the field
  • Rendered documentation · tracks main, which can carry text no tag yet contains
How to cite
APA 7th ed.

Perkins, C. (2026, November 14). KTP and agent identity. Kinetic Trust Protocol. https://kinetic-trust-protocol.net/specs/agent-identity
Plain text

Chris Perkins, "KTP and agent identity," Kinetic Trust Protocol, https://kinetic-trust-protocol.net/specs/agent-identity (accessed November 14, 2026).
full specification
rfcs-txt/ktp-identity.txt
Read KTP-Identity