Trustworthy AI Agent Standards Family

Purpose

This page is the navigation and ownership map for the DPF-originated AI agent standards family. It does not add normative controls. The individual standards remain authoritative for their own subjects.

Standards map

Standard Canonical source Question answered Normative owner
Trusted AI Kernel (TAK) trusted-ai-kernel.md May this agent act, under whose authority, through which tools and data, with what oversight and evidence? Runtime harness, authority, action gating, memory, audit, safety, and earned-autonomy enforcement
Global AI Agent Identification and Governance (GAID) GAID.md Who is this agent, which operating profile is active, what claims are advertised, and how can a relying party verify their status? Identity, AIDoc, claims, badges, receipts, lifecycle, and cross-boundary verification
Job-Specific Intelligence profile (TAK-JSI) job-specific-intelligence.md Is this identified operating profile qualified for this job, activity, data scope, and risk context, and what evidence keeps that qualification current? Job definition, qualification scheme, evidence, surveillance, revalidation, and qualification-to-autonomy boundaries
Gated Permissions Process (GPP) — working draft gated-permissions-process.md Which decision, in which owning authority scope, admits this class of tool use for this stage of this work, and is the model of that envelope complete and actually enforced? The Gated Permission binding (scope gate × capability set × work-shape stage × subject × validity), the capability vocabulary, co-occurrence constraints, model completeness checks, and efficacy measures. Owns no runtime control.

The family’s relationship to NIST, ISO/IEC, IEEE, W3C, IETF, OpenID, 1EdTech, and adjacent protocol work is maintained in the informative External Standards Alignment companion. That document is the single source of truth for cross-standard gap analysis, synergy, augmentation boundaries, and venue allocation. The informative Standards Contribution Roadmap owns engagement sequencing, readiness gates, contribution packages, and go/no-go criteria.

Enterprise operating-model bridge

The Portfolio Aligned Agent and Workforce Operating Standard is an adjacent enterprise standard, not a member of the agent-assurance family. It owns where AI coworkers sit in the four portfolios, how business Products and industry value streams relate to DigitalProducts, how human/AI work is allocated, and how the resulting trace and gaps are assessed. It composes this family whenever an AI coworker is realized as both a managed DigitalProduct and an identity-bearing Performer; it does not redefine TAK, GAID, or TAK-JSI controls.

Composition rule

The standards compose around one governed action:

  1. GAID resolves the enduring AI Coworker identity and the versioned operating profile.
  2. GPP resolves the envelope: the satisfied Gated Permissions for the actor’s Workroom and work-shape stage, each traceable to a gate decision in its owning scope.
  3. TAK-JSI determines whether that profile is qualified for the requested job/activity scope.
  4. TAK intersects the principal’s authority, the coworker’s grants, the GPP envelope (the route/workflow term of TAK §7.2), the data constraints, and the applicable qualification/autonomy ceiling at execution time.
  5. GAID binds the resulting action receipt, the binding and gate-decision references, and the current qualification status back to the identifiable subject.

No document widens another document’s authority:

Principle-directed work and MBSE composition

The working-draft amendment under EP-B932453F makes the governing loop explicit: scoped principles and evidence inform a recommendation; TAK separately checks authority and mediates the effect; GAID binds receipts; JSI evaluates the operating composition and its continuing qualification. Hard constraints precede preference ranking. The normative owners are TAK §7.13, GAID §10.2.1 and JSI §8.4.

The informative MBSE composition design connects requirements, work shapes, decisions, controls and outcome evidence using existing modeling approaches. It is an implementation profile proposal, not a new general-purpose modeling language or a claim of full interchange conformance. These amendments and their proposed assertions still require independent review and execution evidence. Existing implementations do not become conformant by publishing the amended text.

Gated permissions and approval fatigue

Per-call approval fails in practice because people stop reading prompts, and blanket approval fails because it removes the decision altogether. GPP places the decision at the class of work: one recorded decision by the owning scope admits a bounded capability set for a work-shape stage, and TAK enforces it on every call. The evidence and market context are in the informative market and thought-leadership landscape.

Adjacent DPF concepts

Concept Relationship to the standards family
AI Coworker DPF’s managed dual-aspect concept: a DigitalProduct owns lifecycle, release, component, and deployment truth; BusinessProductOffering owns commercial terms, OperationalServiceOffering owns service commitments, and CoworkerOffer/CoworkerEngagement own coworker-service terms and acceptance. An AgentSubject/Performer carries a GAID, operating profile, job qualifications, assignments, and runtime authority under TAK
Portfolio Aligned Agent and Workforce Operating Standard Owns enterprise portfolio, Product, industry-flow, work-allocation, human/AI composition, conformance, and gap semantics around the agent-assurance family
WSID Owns profession and craft doctrine: the knowledge, techniques, evidence practices, and decision axes a job requires
JSI Composes job requirements, WSID material, data/tool/model constraints, evaluations, and outcome evidence into a qualification lifecycle
Proactivity Expresses how readily a coworker should initiate or continue work; it cannot widen authority, qualification, or a regulatory/data ceiling
Earned autonomy The runtime permission level justified by evidence for a specific (coworker × activity × risk) scope
Golden Triangle Compiles human cost/quality/time posture into effort, model tier, verification depth, review depth, and retries; it allocates assurance resources but does not prove competence
Data stewardship Establishes the classification, quality, provenance, use, retention, residency, and accountable ownership constraints that both qualification and runtime routing must honor

Word editions

Each standard and the white paper are also published as a Word document, generated from the Markdown above by pnpm docs:agent-standards. The Markdown is the source of truth; the Word file is a publication snapshot and carries its generation date.

Document Word edition
Trusted AI agent governance white paper Trusted-AI-Agent-Governance-White-Paper.docx
TAK — Trusted AI Kernel Trusted-AI-Kernel-Architecture.docx
GAID — Global AI Agent Identification GAID.docx
TAK-JSI — Job-Specific Intelligence Job-Specific-Intelligence.docx
GPP — Gated Permissions Process Gated-Permissions-Process.docx

Source-of-truth rule

Normative requirements belong in exactly one standard:

White papers, conformance rubrics, diagrams, DPF assessments, and generated Word files are derived companions. They may summarize the standards but must link back to the canonical normative source.

External crosswalks are derived companions and belong in agent-standards-external-alignment.md. Submission execution guidance belongs in agent-standards-contribution-roadmap.md. Specific technical requirements remain with the normative owner named above.

Edit this page on GitHub