The word skill is useful because it describes how an agent learns to do a task. It is not a good boundary for a domain system. A skill is usually a document, a prompt, or a small tool recipe. It can tell an agent what to try, but it does not by itself define the types in a domain, the relationships between them, the evidence behind a result, or the permissions required to change state.

AOE starts one layer below that question. It is an ontology engine: a runtime for a domain model and its corpus. A Domain Package can still ship an Agent Skill as a human-friendly entry point, but the skill is a consumer of the package, not the package's schema or security boundary.

The boundary in one picture

Model Package       Corpus Package       Adapter / Tool Package
(types, relations)  (units, sources)     (providers, validators)
        \                 |                  /
         \                |                 /
          └──────── AOE compiler ─────────┘
                         │
               verified, versioned snapshot
                         │
                 SDK · MCP · HTTP
                         │
                     Agent / app

The engine owns the contracts and the verification rules. The external packages own the vocabulary and data. This is why a new domain should be installable without editing the engine source.

What the engine adds

A model instead of a closed list of kinds. The core has a small meta-schema for declaring types, fields, relations, projections, retrieval, functions, actions, policies, and migrations. The actual domain types are package data.

A corpus instead of a prompt dump. Source units are compiled into an immutable snapshot with deterministic projections, an index, and a manifest. Runtime checks the content digest before serving it, so a changed file cannot silently masquerade as the published version.

A plan instead of a best guess. Query returns a Selection Plan: candidates, feature contributions, hard and soft constraints, relation closure, load order, and token budget. An application can inspect why a unit was selected.

A governed write path. Reading a unit never grants permission to mutate anything. Actions declare their input/output, required capabilities and preconditions. Policy, approval, idempotency, retry bounds, and event evidence are explicit parts of execution.

Where a Skill belongs

An Agent Skill is a useful adapter at the edge. It can teach an agent how to author a Model Package, how to compile a Corpus Package, or how to call a domain's MCP tools. It can also provide examples and conventions. It should not hard-code the engine's ontology or bypass the runtime's checks.

That separation keeps the core reusable across design, support, compliance, operations, or a domain we have never seen. The Frontend Design repository in this workspace is intentionally just one reference Domain Package.

The practical test

If replacing Ticket with Recipe requires changing a switch statement in Core, the boundary is wrong. If replacing it only requires a different Model Package, a different Corpus Package, and perhaps a domain adapter, the engine is doing its job.

For the implementation details, start with the package model, then read compilation and snapshots and selection and execution.