Diop Daily #144 — October 2026

A Tool Call Has Geography

An agent’s tool call is often represented as a name, a schema, and a set of arguments. That description is useful for discovery, but it leaves out the fact that every call occurs somewhere. A request reads data from a jurisdiction, crosses a network boundary, is interpreted by a runtime, touches a system of record, and produces an effect under a particular authority. Geography is already inside the action, even when the API hides it.

This matters because institutions do not govern abstractions. They govern records, facilities, people, contracts, languages, and public obligations located in the world. A support agent that reads a customer file, a research agent that retrieves a restricted corpus, or a procurement agent that submits an order is making a geographic choice as well as a computational one. The institution needs to know where the call was resolved, where its evidence was evaluated, and where its consequence became real.

A machine becomes answerable when it can say where it acted, under whose authority, and which boundary made the action permissible.

Where the request lives

Google’s October 14 announcement of MCP support in Apigee hybrid makes the placement question concrete. In that architecture, the MCP data plane runs inside the customer’s Kubernetes cluster. Tool discovery is generated from an OpenAPI specification, while agent requests are authorized and proxied to the backend inside the customer’s infrastructure. Google states that MCP request and response traffic stays inside the cluster at request time and that configuration is pulled from the control plane rather than pushed into the network.

The significance is architectural rather than promotional. The same tool semantics can be exposed through a managed control plane while the request path remains close to the systems of record. The placement supports data-residency constraints, multi-cloud estates, on-premises systems, and facilities where inbound connections are prohibited. A tool call therefore has at least two locations: the place where its policy and configuration are authored, and the place where its data and effect are handled.

An agent platform that records only the endpoint has recorded too little. A useful execution record should include:

  • Data locality: the jurisdiction, facility, network zone, and storage class from which the input was read.
  • Decision locality: the runtime and policy version that evaluated the request, including the language and identity context used in that decision.
  • Effect locality: the system of record, physical site, account, or public service where the action became consequential.
  • Authority: the principal, delegation, legal basis, policy owner, and expiry that permitted the call.
  • Boundary evidence: traces, attestations, receipts, and failure states showing whether the request stayed inside its declared perimeter.

These fields turn “the agent called a tool” into a statement an institution can examine. They also make the call portable. A second provider can receive the same task while respecting a different approved location, language, or legal basis. Portability becomes the ability to carry the obligation with the work, not merely the ability to point a model at another endpoint.

Jurisdiction is a runtime fact

Data residency is often treated as a storage setting. Agentic systems expose its wider meaning. A record may remain in one country while a prompt, retrieved passage, policy decision, embedding, trace, or generated response crosses into another. The location of the database does not describe the full geography of the action. A proper boundary must cover the path of interpretation and effect, not only the resting place of the source file.

This is especially important for African institutions. A university may hold language data under community conditions that differ from the terms of an external model provider. A public agency may need a case record to remain within a national administrative system. A publisher may have the right to translate a work into Wolof, French, or English while retaining separate obligations for the source corpus and the translated edition. A cooperative may operate across borders while keeping member records under a local custodian. These are governance facts that a generic cloud region label cannot settle.

Sovereignty here is practical. It means that an institution can state which records may leave, which computations must occur locally, which languages are authoritative, which officials may authorize an action, and which evidence must remain available for recourse. A foreign platform may supply a model or a control plane, yet the institution should still own the geographic contract that determines what the machine may touch.

NIST’s AI Agent Standards Initiative places open protocols, agent authentication and identity infrastructure, and security evaluations in one connected programme. That frame is useful because geography cannot be enforced by location alone. The runtime must bind an identity to an authorized perimeter, evaluate the proposed action against policy, and generate evidence that another party can inspect. A request coming from the correct network can still be made by the wrong principal or violate the wrong purpose.

The boundary must understand intent

Google’s September account of zero-trust agents describes a runtime governance layer that screens inputs, evaluates proposed tool calls against user intent and business rules, and watches for anomalous behaviour across a session. The design separates platform governance from the agent’s own code. That separation is important for geography as well. A model should not be the sole interpreter of the boundary it is attempting to cross.

A location-aware gateway can ask questions that an endpoint schema cannot answer: Is this request permitted for this person in this jurisdiction? Is the retrieved record allowed to be summarized by this provider? Does the requested operation change a local system of record or merely prepare a draft? Has the authority expired since the task began? Does the language of the instruction alter the meaning of the policy? Which human or institution receives the case when the location and the purpose disagree?

Intent evaluation also protects local language. A policy written in English may be technically present while its operational meaning is absent for a Wolof, Amharic, Arabic, or French-speaking operator. The control plane should preserve the authoritative policy, its translations, the version used at decision time, and the reviewer who approved the interpretation. Language is part of the geographic boundary because it determines who can inspect, contest, and correct the action.

A serious system should preserve a distinction between three outcomes: the call was allowed in the declared place; the call was redirected to an approved place; or the call was refused because the boundary could not be proved. Silent fallback is a governance failure. If a local runtime is unavailable and the system sends the same record to a distant provider, the institution has experienced a jurisdiction change even if the interface still says “retry.”

Location changes the economics

Execution locality has a cost ledger. Keeping a call near its data can reduce egress, latency, and exposure, but it may require local capacity, specialized operations, and language-specific evaluation. Centralizing a task can simplify maintenance while increasing dependency on an external route, a foreign service window, or a provider’s policy. Local execution can be cheaper or more expensive depending on the boundary; the institution should measure each cost and choose deliberately.

Apigee’s announcement points toward this measurement by describing tool-level quotas, filtered discovery, API performance views, error analysis, latency, target performance, and the ratio of discovery to execution. Those signals can be extended into a geography ledger: time spent at each boundary, data classes crossing each route, egress cost, failed authorization, redirected calls, local cache use, and the evidence retained for each decision. Capital can then examine locality as operating capacity rather than as a compliance slogan.

The same logic applies to infrastructure built for African markets. Intermittent connectivity, mobile-first interfaces, uneven power, multilingual service, and cross-border payment create a different execution map from the one assumed by a continuously connected enterprise. A local control plane may need to queue an action, preserve a signed request, defer a high-consequence effect, and reconcile when the connection returns. The institution’s geography includes the periods when a system is temporarily unreachable.

Where the investable surface is widening

The commercial object is the jurisdiction-aware execution gateway: infrastructure that binds a tool call to its location, data path, language, authority, policy, effect, and evidence before the action becomes consequential.

  • Local execution gateways: MCP, API, and agent runtimes that keep request traffic beside the systems of record while retaining centrally governed configuration.
  • Geographic policy engines: controls that evaluate jurisdiction, residency, purpose, authority, language, and expiry before discovery or execution.
  • Boundary observability: traces and attestations that show where a call was interpreted, which data crossed a perimeter, and where the effect was committed.
  • Portable sovereignty contracts: machine-readable declarations that let an institution move a task between providers without losing locality, delegation, retention, or recourse conditions.
  • Local-language conformance labs: test environments that evaluate whether policies, tool descriptions, and refusal paths preserve meaning across the languages used by the institution.

Underwriting can begin with measurable questions: what percentage of calls carry a verified execution location; how often does a task cross a boundary without an explicit decision; can the institution prove that sensitive records remained in the approved zone; how quickly can it redirect work when a provider or network becomes unavailable; which local-language cases produce different authorization outcomes; and what does a failed locality check cost in delay, egress, human review, or lost service?

The durable asset is the institution’s map of boundaries and exceptions. A gateway becomes more useful as it accumulates tested policies, locality evidence, language cases, provider routes, and recovery procedures. That knowledge can support multiple models and clouds because it describes the institution’s obligations instead of one vendor’s interface. The rail is valuable precisely because it remains when the model changes.

Draw the map before giving the agent a road

An institution can begin with one consequential tool call: reading a protected record, issuing a payment, publishing a translation, or changing a public registry. Draw its full path before automating it.

  1. Name the data, principal, language, jurisdiction, system of record, and intended effect.
  2. Declare the permitted execution zones and the conditions for an approved redirect.
  3. Evaluate the action at the boundary with a policy version that a reviewer can inspect.
  4. Record the trace, authorization, locality, result, and uncertainty without retaining more sensitive content than the purpose permits.
  5. Test the refusal path by removing connectivity, changing the provider, using a language variant, and revoking the authority mid-task.

Repeat the exercise with a local custodian and with an external provider. Compare the evidence, the cost, the delay, and the correction path. The goal is not to make every operation local. The goal is to make every transfer visible, authorized, and reversible where the institution’s obligations require it.

The agentic future will be described through endpoints, models, and tools. Institutions will experience it through boundaries. A tool call that knows where it is, whose authority it carries, and what proof it must leave behind can cross those boundaries without dissolving the institution that owns the action.

Sources