Diop Daily #131 — October 2026

Law Has to Become a Runtime

Artificial intelligence is moving from an object people inspect to an actor that can select tools, touch records, move money, publish material, and change the state of an institution. A legal obligation written in prose does not automatically survive that transition. The rule has to reach the place where an agent chooses, calls, renders, stores, and escalates.

This is the practical meaning of the current governance movement. The European Union’s AI Act is now applicable in stages, with transparency rules taking effect in August 2026 and strict obligations for high-risk systems scheduled for later dates. The Act names risk assessment, logging, documentation, human oversight, robustness, cybersecurity, and accuracy. The Commission also says that technical operationalisation will depend on standards and codes that can adapt to use cases. Law supplies the duty; infrastructure must make the duty executable.

A policy that cannot be evaluated at the moment of action is a wish addressed to a machine.

A rule that cannot travel cannot govern

Traditional compliance often assumes a stable application boundary. A team documents a process, trains staff, reviews access, and samples outcomes. Agent systems disturb that assumption. One workflow may cross a model provider, a tool gateway, a browser, a payment service, a retrieval store, and another agent before a human sees the result. The relevant rule must travel across those boundaries without losing its subject, jurisdiction, purpose, exception, or evidence requirement.

Consider a simple instruction: “Do not disclose a person’s health information outside the care team.” A runtime must resolve several questions before it can enforce the sentence. Which data counts as health information? Who belongs to the care team? Which identity is the agent using? Is the request for treatment, billing, research, or convenience? What happens in an emergency? Which event must be logged? Who can review the decision? A prose policy may answer these questions for a trained professional. An agent needs them as structured conditions and explicit failure paths.

The distinction matters because a model can produce a convincing explanation while the underlying system violates the rule. A prompt may ask for privacy, but the tool may hold broad credentials. A safety instruction may prohibit a transfer, while a connector exposes a second route around the prohibition. A reviewer may approve an output without seeing the hidden context that shaped it. Governance belongs in the execution path, where the system can deny, pause, narrow, route, and record an action.

From legal text to a policy artifact

The first engineering task is to turn a duty into an artifact that can be versioned and tested. The artifact needs more than a yes-or-no rule. It should carry:

  • Subject: which agent, user, institution, data class, tool, or workflow the rule concerns.
  • Purpose: which legitimate task permits the action and which attractive shortcuts remain outside that purpose.
  • Jurisdiction: which law, contract, language community, tenant, geography, or institutional mandate applies.
  • Action: what the system may read, write, infer, publish, purchase, transfer, or refuse.
  • Conditions: the identity, freshness, consent, threshold, human approval, or evidence required before execution.
  • Remedy: how the action is paused, appealed, corrected, reversed, or reported when the conditions are unclear.
  • Evidence: the inputs, policy version, identity, tool call, decision, and human intervention that must remain reconstructable.

This structure turns governance into a portable object. A policy can be attached to a tool, checked at a gateway, carried with a handoff, or evaluated by a separate oversight service. It can also be compared across versions. When a regulation, contract, language rule, or institutional decision changes, the organization can identify which agents and workflows inherit the change instead of relying on a general reminder sent to every operator.

The W3C Agent-to-Web Framework Community Group is evidence that this need has reached the protocol boundary. Its stated mission is to let website operators publish machine-readable governance policies for agent actions, authentication requirements, and human-in-the-loop verification. The group explicitly distinguishes agent action from ordinary crawling and content use. That distinction is commercially important: the web needs a way to say what an agent may do, under which conditions, and with what human involvement.

The runtime is a constitutional layer

Policy enforcement should be distributed across the system rather than entrusted to a single model instruction. Google’s Agent Development Kit documentation lists identity and authorization, input and output guardrails, sandboxed execution, evaluation, tracing, and network controls as separate safety mechanisms. It also recommends deterministic tool restrictions and callbacks that check parameters against agent state or external policies.

Those mechanisms form a constitutional layer for machine action. The model can propose. The runtime decides what the proposal is allowed to touch. The tool can receive a constrained context rather than a vague instruction. The evaluator can test the result against representative cases. The trace can preserve the path by which the action occurred. A human can intervene when the system reaches a boundary that policy cannot resolve automatically.

The constitutional metaphor has a precise use here. A constitution distributes authority, defines limits, and provides procedures for conflict. A policy runtime does the same for an agent fleet. It determines which principal is represented, which capability is delegated, which jurisdiction controls, which evidence is required, and which institution can interrupt the action. The point is not to make the machine behave like a citizen. The point is to prevent capability from becoming authority by accident.

Conformance therefore has to cover behavior, configuration, and evidence. A company can claim that its model follows a policy while its tool permissions contradict the claim. A gateway can block one route while an agent finds another connector. A dashboard can report compliance while omitting rejected actions and ambiguous cases. A meaningful test suite must exercise the whole path, including refusal, escalation, translation, expired consent, stale credentials, policy conflict, and recovery after a mistaken decision.

African institutions need portable rules of their own

Governance infrastructure is also an intellectual sovereignty question. If an institution imports an agent runtime whose policy model is fixed elsewhere, it may inherit categories that do not describe its authority, language, household, archive, or public obligations. The result can look locally translated while remaining institutionally foreign.

A bank, publisher, clinic, university, cooperative, or public office should be able to define its own policy vocabulary and carry it across providers. That vocabulary may distinguish communal custodianship from individual ownership, oral consent from a checkbox, a family relationship from a single account, or a local-language appeal from a generic support ticket. The rule must be machine-readable without becoming conceptually smaller.

This is where African builders can contribute beyond localization. A local policy pack can encode the authority relationships that imported systems tend to flatten. It can specify which language must appear before a consequential action, who may review a disputed decision, what evidence a community archive may expose, and which permissions expire when a project or relationship ends. Such infrastructure gives institutions control over the categories through which agents recognize them.

Where the investable surface is widening

The commercial object is the policy runtime: a layer that compiles legal, contractual, institutional, and linguistic duties into controls an agent system can evaluate, enforce, explain, and prove.

  • Policy compilers: tools that translate governance requirements into executable conditions, approval states, refusal paths, and evidence schemas without pretending that legal interpretation can be fully automated.
  • Agent policy gateways: enforcement points that bind identity, purpose, data class, tool capability, jurisdiction, and human approval before an action reaches an external system.
  • Conformance suites: reusable adversarial tests for prompt injection, privilege drift, policy conflicts, stale consent, translation loss, unsafe retries, and unlogged refusals.
  • Evidence ledgers: systems that connect the policy version, principal, context, tool call, decision, intervention, and outcome so an institution can demonstrate what happened.
  • Jurisdiction and language packs: maintained rule libraries that preserve local legal concepts, institutional categories, terminology, and routes of recourse across models and providers.

Underwriting can begin with measurable capacity. Can the organization identify every action governed by a policy? How quickly can it revoke a capability across agents and providers? What proportion of consequential actions carry a complete evidence record? How often do policies disagree across tools? Can a reviewer reproduce the same decision from the stored policy version and inputs? Can the institution export its rules, tests, and evidence when it changes infrastructure?

The strongest companies in this layer will sell operational certainty rather than a promise that a model is safe. Their customers will include regulated enterprises, public bodies, platforms, model providers, and institutions operating across jurisdictions. The durable asset is a maintained corpus of policies, tests, mappings, exceptions, and evidence. That corpus improves with use because each ambiguous case can become a clearer rule or a better escalation path.

Build one policy boundary first

An institution can begin with one workflow that creates a real consequence: a payment, a benefits decision, a health-data transfer, a publication, or a change to a production system. Map the actors, data, tools, jurisdictions, approvals, and failure states. Write the policy as a structured artifact. Enforce it at the tool boundary and at the human approval surface. Generate a record for both allowed and refused actions.

Then test the boundary in the languages and operating conditions that matter. Alter the identity. Remove the source. Expire consent. Introduce a conflicting policy. Ask the agent to use a permitted tool for an impermissible purpose. Disconnect the primary provider. Require a human appeal. A policy is becoming infrastructure when these tests produce bounded behavior and evidence instead of a confident paragraph.

The legal order will continue to change. Models will change faster. Institutions that wait for a final rule or a final model will remain exposed to the gap between what they are required to do and what their systems can prove. The builders who close that gap will give organizations something more durable than compliance theatre: a runtime in which authority, limits, language, evidence, and correction can travel together.

Sources