Ovia Inc — Artificial Intelligence

Context Engineering

Semantic layers, ontologies, and context graphs — how shared meaning and personal permissions decide whether agents act on the truth.

Meaning is infrastructure

Every organisation runs on definitions. Revenue, customer, premier, active, late: each word carries a rule, and each team quietly keeps a private version of the rule. Dashboards disagree, reports need reconciliation meetings, and agents trained on the disagreement produce fluent nonsense. Context engineering begins by treating meaning as infrastructure: definitions written down, versioned, governed, and shared by every person, dashboard, and agent that touches the business. The work sounds unglamorous next to model announcements. The work is also the difference between a demonstration and a system. A model can only act on the context supplied; when the context contains three versions of the truth, the model picks one silently and the error looks like confidence.

The semantic layer learns the business

A semantic layer translates between business language and data. Early versions held definitions: this column means revenue, that filter means this region. The newer generation goes further, turning business rules into executable, versioned semantics: a phrase such as Premier Customer becomes a testable definition that systems can apply identically everywhere. Subagents now specialise in exactly that translation work, interviewing the business through its data and proposing rules for humans to approve. The ordering matters. After definitions come rules; after rules, judgment. A metric computed correctly is only step one. A metric whose definition the whole company shares, whose rule changes are reviewed and versioned, becomes a fact the company can act on without a meeting.

One ontology, many models

Meaning drifts wherever models multiply. Each new dashboard, warehouse, and agent adds a local dialect, until the same word names different things in different rooms. The answer is not to freeze every team into one database. The answer is one governed ontology: a single, maintained statement of what the core concepts are and how the concepts relate, mapped onto the many semantic models that already exist. Each system keeps a local shape; the ontology keeps the translation honest. When an agent moves between systems, the agent carries the shared meaning along, the way a good translator carries intent rather than words. Governance sounds slow. Drift is slower: drift costs a reconciliation meeting for every decision, forever.

Context travels, authority stays personal

Context graphs add the next layer: not just definitions, but relationships, history, and the trail of how knowledge connects. Platforms now place those graphs inside the data warehouse itself, where agents can read across them. The critical design decision is permission. In the pattern that works, an agent reading a context graph inherits the permissions of the person asking, nothing more. Context travels to wherever the question is asked; authority stays personal, attached to a named human with a defined right to see. The companion discipline comes from zero trust thinking applied to agents: each agent carries an identity of its own, permissions stay minimal, and the record of what was accessed remains verifiable. An agent with broad silent access is a liability wearing a productivity costume.

Where context work pays off

The returns arrive in ordinary places. A support agent answers with the definition the business actually uses, so the answer matches the invoice. A forecasting agent reads the same ontology as the finance dashboard, so the argument in the meeting is about the future instead of about the numbers. An operations agent searches a context graph under the permissions of the engineer on call, fast and bounded at once. None of these wins requires a new model. Each win requires the unglamorous labour of writing meaning down, versioning the writing, and wiring access to named people. That labour is the core of context engineering, and the reason the discipline sits at the centre of our practice: models change every quarter, while meaning, once governed, compounds.

Starting without a programme

Context work does not need a platform purchase or a year of committee time to begin. A starting pass fits in weeks. Choose the ten definitions the business argues about most, the words that trigger reconciliation meetings, and write each one down with an owner, a rule, and a date. Publish the ten where every dashboard builder and agent designer will trip over them. Connect one agent or one report to those definitions and measure what changes: fewer corrections, shorter meetings, answers that match the invoices. Then extend the list, twenty definitions, then fifty, versioning each change as the business learns. Governance grows by evidence, not by decree. The ontology that governs a whole estate begins as a short list that a small team defended successfully, and the credibility earned by the first ten definitions carries the next hundred.