Diop Daily #127 — September 2026

Before the Call, Give the Institution an Address

An autonomous agent reaches a web service before it reaches the institution behind that service. It encounters a domain, an endpoint, a certificate, a tool list, a schema, and a response. None of those objects, by themselves, tell the machine what the resource is for, who is responsible for it, which actions are outside its jurisdiction, or whether the context it has inferred is reliable.

That gap is becoming a market and a constitutional problem. W3C’s Introduction Layer Community Group, launched in August, describes the missing primitive directly: a web resource should be able to introduce its identity, purpose, constraints, and responsible party to an AI agent in a discoverable, machine-native form. The proposal is still a community initiative, not a settled standard. Its importance lies in the question it forces into the open: what must an institution say about itself before a machine is allowed to act around it?

A machine cannot respect a boundary that the institution has never described.

An endpoint is a door, not a public identity

Software architecture has spent decades making services callable. OpenAPI describes operations and data shapes. JSON-LD describes structured meaning. Certificates establish control over a domain. Access tokens establish that a principal may enter a protected route. These pieces are necessary, yet they answer different questions.

An institution also needs to state its purpose, its scope, its accountable owner, its operating jurisdiction, its constraints, and the circumstances in which it will refuse a request. A public archive may expose a search endpoint while prohibiting bulk extraction of living-person records. A cooperative may offer inventory data while requiring member consent before an order is settled. A cultural publisher may authorize quotation while reserving translation, adaptation, and commercial reuse. The endpoint can be technically healthy while the machine-facing meaning remains absent.

W3C’s Introduction Layer proposal distinguishes this problem from capability discovery. WebMCP may declare what an agent can do. An introduction layer would explain what a resource is, what it is responsible for, and what must remain outside the interaction. That ordering matters. An agent should learn the nature and jurisdiction of a resource before it selects the operation that appears most useful.

The first document an institution gives a machine therefore resembles an address more than a manual. An address locates a person or organization within a social and legal space. A machine-readable address should carry at least:

  • Identity: the resource, organization, or authority that is speaking.
  • Purpose: the public function the resource exists to perform.
  • Responsibility: the party that maintains the resource and answers for its representations.
  • Constraints: actions, data uses, jurisdictions, and categories that are prohibited or conditional.
  • Evidence: the records, credentials, and update history that support the introduction.
  • Recourse: the route through which a person or institution can challenge a machine-mediated interpretation.

Identity is a boundary with a history

NIST’s AI Agent Standards Initiative places agent authentication and identity infrastructure alongside open protocols and security evaluations. That arrangement is instructive. Identity lets a receiving system ask who is acting, under whose authority, and with what evidence; interoperability depends on that identity.

The W3C Holon Community Group offers a useful systems vocabulary for this work. Its September update describes entities as systems that contain and connect to other systems, change over time, expose projections of their internal state, and operate under constraints. The analogy to institutions is precise enough to be useful: a university, archive, ministry, cooperative, or laboratory cannot expose its entire interior to every requester. It exposes a surface through which an external actor infers the state, authority, and limits of the whole.

That surface needs versioning. A resource may change ownership, jurisdiction, policy, language coverage, or data retention rules while keeping the same domain. A machine that cached the old introduction can therefore make a current decision with obsolete premises. The introduction layer must carry an update time, an effective period, a version, and a way to retrieve the authoritative revision. Temporal validity belongs in the identity record because a boundary that expired yesterday is a different boundary today.

This is also where African sovereignty becomes concrete. Many African institutions are asked to present themselves to imported platforms through categories designed elsewhere: organization type, address format, language, legal status, beneficiary, customer, or risk. A machine-readable address gives the institution a place to declare its own distinctions before an external agent compresses them into a foreign schema.

A Senegalese archive may need to distinguish custodianship from ownership, oral permission from written permission, and a French catalogue label from the name used by the community. A pan-African payment cooperative may need to state the relationship between a member, a household, a local agent, and the account that settles a transaction. These are semantic and constitutional choices. They cannot be delegated to a model’s prior assumptions without surrendering part of the institution’s authority.

Provenance turns an introduction into a claim

An institution’s address is useful only when a machine can tell the difference between a declaration and an attested fact. W3C’s AI Computed Provenance Community Group, launched in August, frames this distinction as “Computed ≠ Asserted.” Its proposed work concerns cryptographic provenance for AI-assisted science and technology, along with conformance tests that allow an unrelated party to reproduce a result.

The same discipline applies to institutional introduction. A resource may claim that it represents a ministry, archive, publisher, or cooperative. The machine needs a way to inspect the binding between that claim and the authority that controls the resource. It also needs to know which parts are computed from current records, which are declared by an administrator, which were signed by an external authority, and which remain unverified.

W3C’s Agent Conformance and Benchmarking Community Group sharpens the operational requirement. Its August proposal calls for weighted rubrics whose checks name observable evidence, runnable tests with adversarial fixtures, versioned corpora, and published results that an unrelated party can reproduce. An introduction layer without conformance would become another field of optimistic metadata. Conformance gives the address a testable relationship to the resource it describes.

A serious boundary profile could therefore expose a small evidence ledger:

  1. the identity claim and its issuer;
  2. the purpose and jurisdiction that the institution has declared;
  3. the constraints that were tested against the live interface;
  4. the current policy and data-retention version;
  5. the provenance of important responses and changes;
  6. the reproducible checks that support the profile’s conformance score.

This does not create certainty. It creates a better basis for judgment. A machine can proceed with a bounded confidence, pause for a human, or refuse to cross a boundary when the evidence is missing or stale. Trust becomes a sequence of inspectable conditions rather than a mood generated by a polished interface.

The sovereign interface begins before the API

Builders often begin with the action they want an agent to perform. Institutions should begin one step earlier with the context in which that action could be legitimate. This changes the implementation order:

  • write the institution’s purpose and responsible party in machine-readable form;
  • declare the jurisdictions, languages, populations, and data categories that shape interpretation;
  • separate public description from privileged discovery and invocation;
  • attach current policy versions, evidence links, and update conditions;
  • test whether independent agents infer the same boundary from the profile;
  • publish a human route for correction, appeal, and institutional contact.

The language step is especially important. A French introduction that loses a distinction present in the English version is an identity defect, not a cosmetic translation issue. A local-language name that cannot be discovered is a distribution defect. A refusal translated into a success-shaped phrase is an authority defect. Machine-facing language belongs inside the institution’s quality and governance process.

This approach also protects small institutions. A university department, community archive, cultural cooperative, or municipal service does not need the infrastructure budget of a global platform to publish an accountable machine address. It needs a compact profile, a stable identifier, a signing or verification path, a conformance test, and a human route for disputes. The technical surface can be small while the constitutional content remains serious.

Where capital should look

The investable object is the institutional introduction layer: infrastructure that makes a resource recognizable, bounded, current, and answerable before an agent invokes its capabilities. It is distinct from a tool registry. A tool registry catalogs actions. An introduction layer establishes the authority and context in which those actions can be understood.

  • Boundary-profile registries: versioned records for identity, purpose, responsibility, jurisdiction, constraints, languages, and recourse.
  • Identity and authority envelopes: credentials that bind an agent, service, or organization to the claims it presents across domains.
  • Conformance runners: reproducible tests that compare an introduction with live behavior, policy, access control, and response evidence.
  • Provenance adapters: systems that attach issuer, computation, transformation, and update history to machine-mediated outputs.
  • Multilingual boundary services: translation and terminology layers that preserve jurisdictional meaning across French, English, African languages, and domain vocabularies.
  • Recourse infrastructure: logs, contact routes, appeal queues, and correction workflows for people affected by an agent’s interpretation.

Underwriting can start with observable measures: the rate at which independent agents infer the same purpose, stale-profile detection time, unauthorized discovery attempts, conformance reproducibility, policy-change propagation, language-specific interpretation error, and time from a disputed machine action to a human response. These measures describe whether an institution can be encountered without being misrepresented.

The strongest products in this category will sit between standards and operations. They will turn a specification into a maintained profile, a profile into a testable boundary, and a test result into evidence a buyer, regulator, partner, or community can inspect. That is a more durable position than selling another opaque agent wrapper.

Give the machine a place to stand

The next generation of agent systems will encounter more than APIs. They will encounter archives, schools, publishers, hospitals, public offices, cooperatives, and research laboratories whose authority cannot be reduced to a list of callable functions. Each resource will need to announce where it stands, what it serves, what it refuses, who answers for it, and how its claims can be checked.

Africa should participate in this layer as a producer of institutional vocabulary, not only as a consumer of imported interfaces. The continent’s languages, legal pluralities, payment systems, archives, and community structures contain distinctions that global machine schemas routinely flatten. Publishing those distinctions in durable, testable forms is technical work and political work at once.

Before a machine calls the institution, the institution should have the right to introduce itself. That first address is where recognition becomes authority, and where an interface can begin to serve sovereignty rather than silently replacing it.

Sources