Anthropic Skills Lacks Architectural Cross-Authority Substrate

Vendor & Product Reality: Anthropic Skills as Capability Substrate

Anthropic Skills, introduced as a packaging format for Claude-invocable capabilities, provides a structured way to deliver tool schemas, instruction text, reference materials, and execution scaffolding to a Claude session. A skill is essentially a declarative bundle: a manifest describing the skill's purpose, the tools it exposes, the prompts that orient the model toward those tools, and any auxiliary files the model may consult. At inference time, the platform routes the skill into context based on user invocation, agent self-selection, or platform-curated routing rules. Anthropic supplies first-party skills, customers author private skills, and a marketplace-like distribution model is emerging for shared third-party skills.

The technical execution at platform scale is mature. Skill discovery, loading, token-budget management, tool dispatch, and result integration all operate at the latency and reliability levels Claude users expect. For single-tenant, single-jurisdiction enterprise workflows — a U.S. customer using U.S.-hosted Claude with U.S.-authored skills under one customer admin domain — Skills functions as designed and represents the current state of the art in agentic capability extension.

The product is positioned as the durable substrate for agent capability across Anthropic's product line: Claude.ai, the API, Claude Code, Bedrock and Vertex deployments, and forthcoming agent-platform releases all consume the same skill format. That uniformity is a product virtue — and also the reason the architectural gap below matters at platform scope rather than only at the edge.

Architectural Gap: Cooperative Loading Without Cryptographic Authority

Skill loading today is cooperative. The platform admits a skill into a session because the platform's policy engine says it should be admitted; the skill carries metadata describing its origin, but that metadata is not cryptographically bound to the runtime invocation in a way that survives composition. When a single agent session loads skills authored by Anthropic, by the deploying enterprise, by a third-party vendor, and by a regulated-domain authority simultaneously, there is no structural mechanism that enforces "this tool call was authorized by the intersection of all four authorities under the operating jurisdiction's rules." The session executes; logs are produced; an after-the-fact audit can attempt reconstruction. Authority is observable, not constitutive.

Cross-jurisdiction agent operations expose the gap. The EU AI Act's high-risk system requirements, U.S. AI Executive Order obligations on dual-use models, Japan's METI agent guidelines, and emerging sector-specific agent regulations in finance and healthcare each impose structural constraints on what a skill may do, where it may execute, who may have authored it, and which composite authority must consent before action. A skill admissible in one jurisdiction may be inadmissible in another; a skill admissible to one customer's data domain may be inadmissible to a partner's. Cooperative loading cannot encode those constraints as runtime preconditions; it can only encode them as platform-side gating that breaks the moment a skill is replicated, forked, sideloaded, or composed into a derived capability.

The deeper architectural problem is that the skill format treats authority as descriptive metadata rather than as a load-bearing element of the invocation contract. There is no cryptographic primitive that says "this skill, executed in this environment, against this data, by this operator, has the joint signatures required to act." Without that primitive, every regulatory regime that demands verifiable cross-authority consent must be approximated through platform policy — and platform policy is a single point of governance that does not federate.

What the Spatial-Adaptation Primitive Provides

The spatial-adaptation primitive treats each skill activation as a bounded, environment-conditioned, cryptographically-signed event. A skill is not merely loaded; it is admitted into a defined operational space whose boundaries are declared by the composition of authorities that signed for the activation. The primitive supplies three structural elements that the current Skills format lacks: a runtime-signed adaptation envelope that binds the skill's authority chain to the specific invocation context; a composite admissibility evaluation that requires every relevant authority — platform, customer, regulator, third-party author — to have signed for the activation under the present environmental conditions; and a federation contract that allows authority chains to compose without requiring any single authority to be the universal trust root.

Concretely, when a Claude session running in an EU-deployed enterprise tenant attempts to invoke a clinical-summarization skill authored by a U.S. healthcare vendor, the primitive evaluates whether the vendor's authoring signature, the enterprise customer's deployment signature, the EU AI Act admissibility signature for the operational context, and the platform's runtime signature all compose into a valid envelope for this exact invocation. If any signature is missing or any environmental precondition is unsatisfied, the skill does not execute — not because a policy engine blocked it, but because the envelope is structurally invalid. The cryptographic binding makes the authority composition a precondition of execution rather than a post-hoc audit artifact.

Equally important, the primitive externalizes the authority substrate. Anthropic remains a credentialed authority — typically the runtime-signing authority and often the curating authority — but ceases to be the only path through which cross-jurisdiction skill composition can occur. Customers, regulators, and third-party authors can each operate their own signing infrastructure; the primitive specifies how their signatures compose without requiring them to terminate at a single root.

Composition Pathway: Layering the Primitive Above Existing Skills

The composition is additive, not replacement. Anthropic's existing skill format, marketplace, routing, and execution machinery continue unchanged. The spatial-adaptation primitive sits above the skill loader as a precondition layer: when the loader prepares to admit a skill into a session, the primitive evaluates the composite envelope and either issues a runtime-signed admission token or denies admission with a structured reason chain. Existing skills that operate within a single authority domain see no behavioral change; the envelope evaluation produces a trivially valid token because there is only one authority to consult.

For multi-authority deployments — the cases where the gap currently bites — the primitive provides a declarative specification language for expressing which authorities must sign, under what environmental conditions, with what scope of action, and for what duration. Enterprise customers express their composite admissibility once; regulators publish their jurisdictional admissibility specifications; third-party skill authors declare their authoring authority and the scopes they assert. The primitive composes these declarations at activation time. The activation either satisfies the composite or it does not, and the result is a verifiable runtime artifact rather than a logged inference about platform behavior.

Migration from cooperative loading to cryptographic envelopes can be staged. Initially, the primitive can operate in observe-only mode, producing envelope tokens alongside normal skill execution so that customers can validate their authority specifications against actual workload patterns. Subsequently, the primitive enters enforce mode for skills marked as requiring composite admissibility, with the marking driven by regulatory scope, customer policy, or skill-author declaration. Finally, federation-mode operation allows authority chains to extend across platforms — a skill admitted under a composite envelope on Claude can present the same envelope token to a downstream tool or external service, and the downstream actor can verify the envelope without contacting the original platform.

Commercial & Licensing Trajectory

Anthropic gains the architectural cross-authority adaptation layer above Skills, which converts the Skills product from a single-tenant capability framework into a federation-ready substrate suitable for the regulatory environment that EU, U.S., and APAC agent rules are converging toward. Cross-jurisdiction agent operations gain structural support; regulatory exposure gains structurally-verifiable cross-jurisdiction compliance instead of best-effort platform-policy approximations; multi-customer enterprise agent operations gain governance that survives skill replication and federation rather than depending on a single platform's gating.

The patent positions the spatial-adaptation primitive at exactly the architectural location where cross-jurisdiction agent evolution is forcing the question. Anthropic's competitive position benefits from adopting the primitive as a licensed substrate: the Skills product retains its market leadership while the substrate provides the verifiable authority composition that enterprise and regulated-sector customers will increasingly require as a condition of adoption. Licensing the spatial-adaptation primitive into the Skills runtime is the lowest-friction path for Anthropic to deliver verifiable cross-authority skill composition without rebuilding the format, the marketplace, or the developer surface.