Diop Daily #141 — October 2026

The Package Must Explain Itself

An agent that can install a capability has crossed a threshold that a conventional software installer never had to face. The machine is not merely fetching code. It is deciding whether a bundle is relevant, what authority it requires, which dependencies it can touch, how it should be evaluated, and whether its owner has made a claim that the institution can accept. A package that cannot answer those questions remains opaque at the moment it becomes operational.

This changes the unit of distribution. The useful object is not a repository with a README added for humans after implementation. It is a capability package whose identity, purpose, inputs, permissions, dependencies, tests, ownership, and removal path are legible before installation. The package must explain itself because the next operator may be an agent acting on behalf of a person who will never inspect the source tree.

When software can recruit its own operator, metadata becomes part of the control plane.

The manifest arrives before trust

Google’s October 15 description of google/adk-recipes makes this shift unusually concrete. The repository is organized into reference patterns, installable vertical plugins, and deployable services. Each recipe carries a validated manifest, an AGENTS.md file that instructs a coding assistant how to use it, and, for the deployable layers, a container definition and runnability test. The repository is designed to be discovered and consumed by agents as well as developers.

That structure is a product decision. A manifest is not decorative documentation; it is an admission surface. It tells a machine what a capability claims to do, what it owns, how it is licensed, which environment it expects, and which checks stand behind the claim. The assistant can then compare a candidate package with a task, a policy, a budget, and a security boundary before it writes files or invokes a tool.

A capability package should expose at least:

  • Identity: a stable name, version, publisher, ownership record, and relationship to earlier releases.
  • Function: the task it performs, the inputs it accepts, the outputs it produces, and the cases it refuses.
  • Authority: the files, networks, secrets, accounts, data classes, and external services it may reach.
  • Dependencies: models, runtimes, libraries, tools, registries, and human approvals required for a valid execution.
  • Evidence: tests, evaluation fixtures, supported environments, known limitations, and the date on which the claims were checked.
  • Correction: update, revocation, rollback, deletion, and notification procedures when the package or one of its dependencies changes.

These fields turn discovery into a reasoned decision. They also give the institution a way to refuse a package without pretending that refusal is a failure of intelligence. An agent may reject a useful capability because its license is unclear, its requested permission is too broad, its evaluation is stale, or its correction path is missing. A bounded refusal is an operating feature.

Distribution is a governance event

Software distribution has always carried authority, but the authority was often mediated by a person. A developer read the documentation, inspected the dependency tree, ran a test, and decided whether to merge. Agentic distribution compresses those steps into a runtime that must act quickly while preserving the institution’s right to understand and reverse the decision.

NIST’s AI Agent Standards Initiative places open protocols, authentication and identity infrastructure, and security evaluations in the same strategic frame. The initiative describes work on agents that operate securely on behalf of users and interoperate across systems. Its importance for capability packages is direct: identity tells the institution who is making the claim, evaluation tells it what has been checked, and open protocols determine whether the package can be moved or examined outside one vendor’s environment.

Google’s same-day guidance on evaluating AI agents supplies the operational discipline. It recommends beginning with critical user journeys, keeping a frozen golden dataset while changing one variable, combining human review with automated checks, and measuring latency, token consumption, cost, and tool-call counts alongside answer quality. It also argues for trace conventions that keep evaluation independent of a particular orchestration framework.

The package therefore needs a release record, not only a version number. That record should say which journeys it passed, which environment produced the result, which checks were boolean and repeatable, which limits remain open, and what changes require a new evaluation. A package that silently changes its tools or permissions while preserving its version has not maintained compatibility; it has changed the institution’s risk.

The African question is who owns the catalogue

Agent-ready distribution will produce new catalogues of skills, plugins, recipes, datasets, connectors, and deployable services. The catalogue will influence which capabilities are visible, trusted, affordable, and easy to install. That makes the registry a political and economic institution, not a neutral shelf.

African laboratories, publishers, universities, cooperatives, and public agencies should be able to publish capabilities in their own languages, describe local data categories, set jurisdictional limits, and retain the authority to withdraw or revise a package. A registry that only accepts the categories of an external platform can make local knowledge available while making local institutions invisible. Interoperability should expand the vocabulary of participation, not narrow it to whatever a foreign marketplace can index.

This is where historical method becomes an engineering discipline. A source is not strong because it is present in an archive; it is strong when its origin, context, limits, and chain of transmission can be examined. A capability package deserves the same treatment. Its manifest should let another institution ask where it came from, what it assumes, what it can affect, and how to challenge its claims.

Where the investable surface is widening

The commercial object is the capability distribution layer: registries, manifests, evaluation records, permission boundaries, and lifecycle services that let agents find and install software without turning adoption into an unreviewable act.

  • Capability registries: searchable stores that index function, ownership, permissions, dependencies, language, jurisdiction, and evidence rather than keywords alone.
  • Manifest and signing services: tools that bind package metadata to a publisher, version, release, and integrity proof while preserving portable export.
  • Evaluation admission: systems that run critical user journeys, frozen fixtures, security checks, and cost budgets before a package enters a trusted catalogue.
  • Permission-aware installers: runtimes that compare requested access with institutional policy, isolate execution, and require a new decision when a release changes its boundary.
  • Revocation and migration rails: services that notify dependants, stop unsafe versions, preserve the decision record, and move an institution to an alternative package without losing operating knowledge.

Underwriting can begin with measurable questions: how quickly can a reviewer understand a package before installation; what proportion of capabilities have named owners and explicit permissions; how often do package updates invalidate prior evaluation; can an institution export its catalogue and decision history; how many failed installations are contained before they reach sensitive systems; and can a local publisher make its capability discoverable without surrendering its language or licensing terms?

The strongest companies in this category will not simply host more code. They will reduce the uncertainty between discovery and governed use. Their assets will be trusted metadata, reproducible evaluation, policy integrations, installation telemetry, and a registry that remains useful when a model, framework, or cloud provider changes. The rail earns revenue because it lowers the cost of adoption without hiding the cost of responsibility.

Build one package that can survive handoff

An institution can test the principle with one capability that already crosses teams: a research importer, a translation skill, a data connector, a release assistant, or a publishing workflow. Give it a stable identity. Write the manifest before the installer. State the permitted inputs and outputs, the secrets it may request, the environments it supports, the tests that define success, and the conditions that require refusal.

Then hand the package to a second team using a different runtime. Ask them to discover it, evaluate it, install it in a sandbox, inspect its permissions, run the critical journey, and remove it. Change one dependency and see whether the registry marks the evaluation as stale. Revoke one permission and verify that the package stops at the boundary rather than finding an undocumented route around it.

This is the discipline required for an agentic software economy. Distribution will become faster, but speed without legibility creates dependency on whoever owns the catalogue. A package that explains itself gives an institution a chance to accept, reject, adapt, and carry capability on its own terms. The machine may perform the installation. The institution must still own the decision.

Sources