Diop Daily #126 — September 2026

The Description Is Part of the Tool

A tool that an agent cannot interpret is an inaccessible capability, even when its endpoint is healthy. The description, schema, authentication rule, quota, and response shape determine what the machine believes it may do before a single request is sent. That makes the tool declaration part of the product, part of the security boundary, and part of the institution's operating language.

Recent public signals make this visible. Google's September 24 announcement of remote Model Context Protocol support in Cloud API Gateway says that a tool's description is the primary signal an LLM uses to decide when to call it. The same announcement describes MCP discovery and REST invocation sharing authentication, quota, and logging policies. Although it is a product announcement rather than independent proof of safe deployment, it exposes the architectural seam: a machine meets an institution first through a declared vocabulary, then through an endpoint.

Before a machine can obey an institution, it has to recognize the institution's terms.

The tool is encountered before it is invoked

Human developers often treat documentation as a wrapper around the real software. An agent experiences the order differently. It encounters a name, a description, an input schema, a permission context, and an advertised result. Those elements shape the choice to invoke, the arguments it constructs, and the confidence it assigns to the response.

This turns language into an operational surface. “Search customers” does not carry the same boundary as “search active customer records for accounts assigned to the requesting team, returning only fields authorized by the account policy.” The second declaration gives the machine a jurisdiction, a population, and a limit. It also gives a reviewer something concrete to inspect when the call is wrong.

The distinction matters for small institutions as much as for large platforms. A community archive, a cooperative inventory system, or a public information service may have an API that works perfectly for a human engineer while remaining semantically opaque to an agent. The missing asset is often not another model. It is a vocabulary that states what the operation means, who may use it, what evidence it returns, and what a refusal looks like.

Discovery is already a policy decision

Google's API Gateway announcement contains a quiet warning alongside its convenience. In the public-preview design, tools/list is unauthenticated by default, which makes development easy while publishing tool names and input schemas to anyone who asks. Production systems are told to require a JSON Web Token for discovery, while the underlying REST authentication still governs tools/call.

The lesson is broader than one gateway: discoverability is a form of exposure. A tool can be protected at execution time and still reveal sensitive institutional structure at discovery time. Names, descriptions, fields, and error messages can disclose the shape of a ledger, the existence of a vulnerable workflow, or the categories by which a public office sorts people.

A serious tool contract therefore answers two different questions:

  • Recognition: what may an authorized machine understand about this capability, and under which identity may it discover that description?
  • Invocation: which principal may call it, with which arguments, against which records, under what quota, and with what logged result?

Collapsing these questions produces a thin notion of access control. An agent can be allowed to call a tool while being given a misleading or overbroad description. It can also be denied at the endpoint after learning more than it should about the institution. The boundary needs both semantic precision and policy continuity.

Schemas carry jurisdiction

A schema is often described as a technical shape for data. In agent systems, it also carries a theory of jurisdiction. Field names decide which distinctions are visible. Required fields decide which facts must exist before an action can proceed. Enumerations define the categories the machine can select. Error states decide whether uncertainty becomes a refusal, a retry, or a silent omission.

That is why imported schemas deserve the same scrutiny as imported models. A payment tool built around card settlement may omit the consent and reconciliation steps required by a mobile-money cooperative. A research archive schema may preserve an English title while discarding the local-language name by which a community recognizes the work. A customer record may encode household relations as irrelevant text even when those relations determine who may authorize a transaction.

W3C's June announcement of a workshop on e-commerce for humans and AI agents makes the representation problem explicit. The web has historically been shaped for human readers, even when pages also expose structured data through APIs or embedded JSON-LD. Agents are becoming intermediaries between content and users, so creators and publishers must consider how machine systems interpret what they publish. The workshop focuses on e-commerce, yet the underlying issue reaches archives, public services, education, and cultural production: what a machine can recognize becomes part of how an institution can be reached.

Machine recognition should therefore be treated as an editorial and constitutional task. The institution chooses the categories, examples, language variants, refusal states, and evidence requirements that define a legitimate operation. External frameworks can provide transport and syntax. They cannot decide what a local record means.

Local execution does not remove the need for a contract

Google's September 23 announcement of local model support in the Antigravity SDK describes offline execution, local GPU and RAM use, privacy benefits, and hybrid workflows in which a cloud model plans while local models perform work on-device. The announcement uses a recorded demonstration to report that most tokens in that run were handled locally; that figure remains a vendor-reported example rather than a general performance guarantee.

The architectural point survives the marketing boundary. Local execution can keep code, requests, or culturally sensitive records within an institution's physical or legal perimeter. It does not automatically make the institution's interfaces legible. A local agent can still misunderstand a Wolof term, infer authority from the wrong field, or call a tool whose description hides a consequential side effect.

NIST's AI Agent Standards Initiative places open protocols beside authentication, identity infrastructure, and security evaluations. Its public framing supports a useful division of labor: protocols can help systems interoperate, identity can establish who is acting, and evaluation can test whether the interaction is secure. The institution still needs to define the local semantics that make a call legitimate. Sovereignty lives in that definition, not in the location of the GPU alone.

Where the investable surface is widening

The commercial object is the tool contract: the shared semantic and conformance layer between an institution's capability and the machine that seeks to use it. It sits closer to a revenue-bearing boundary than generic documentation because it determines discovery, eligibility, execution, evidence, and correction.

  • Contract registries: systems that version names, descriptions, schemas, examples, refusal states, owners, and policy links as one reviewable unit.
  • Discovery controls: authorization for who may see a capability, its fields, its jurisdiction, and its machine-readable metadata.
  • Conformance testing: tests that ask whether the declared meaning matches the live endpoint, including permissions, quota behavior, error states, and response provenance.
  • Semantic translation: layers that preserve local terms, multilingual distinctions, and institutional categories when a common protocol crosses language or system boundaries.
  • Boundary observability: records of what an agent believed it was calling, what the tool actually received, which policy applied, and how a human corrected the mismatch.

Underwriting can begin with measurable questions. How often does a description lead to an invalid call? Which fields cause the most human correction? Can a new model discover the same capability without being granted broader access? Do schema changes trigger review before they alter agent behavior? Can the institution export the contract, its test cases, and its decision history when it changes gateway or model provider?

OpenAI's September account of AI-native workflows supplies a complementary market signal. The examples of Basis, Clay, and Exa connect agents to company context and tools, preserve persistent workspaces or reusable skills, and keep tests and human review near consequential actions. The cases are vendor-reported, so they should be read as design evidence rather than universal adoption data. Their common pattern is clear: a workflow becomes repeatable when context, permissions, evidence, and decision rights remain attached to the work.

Build one contract before adding another agent

A practical experiment needs one real tool, one named owner, and one consequential boundary. Write the description as if a new machine will encounter it without a private conversation. Define the records it may touch, the identities it accepts, the fields it refuses, the evidence it returns, and the circumstances that require a person. Test discovery separately from invocation. Then ask an independent reviewer to compare the declaration with the live behavior.

Repeat the test in the languages and institutional settings that the system claims to serve. A French description that loses a distinction present in the English schema is a conformance defect. A translation that makes a refusal sound like success is a safety defect. A local-language name that cannot be searched is a distribution defect. These are engineering failures with social consequences, and they can be measured before a fleet makes them expensive.

The next stage of agent infrastructure will be judged at these boundaries. The winning systems will help institutions expose capabilities without exposing more authority than intended, preserve local meaning across machine handoffs, and prove that the declared contract still matches the action. A machine does not enter an institution through intelligence alone. It enters through the words, fields, permissions, and evidence that tell it where it is.

Sources