Building a Semantic Foundation for Harness Engineering: Ontology‑Driven Controllable Agents

Although Agent technology has become the dominant form of enterprise AI in 2024‑2025, many deployments fail to stay within business boundaries because the agents lack an explicit "rule‑aware" structure; they can act confidently but often make mistakes that violate company policies.

To achieve "safe and controllable execution," the author decomposes the problem into independent dimensions: permission & isolation, behavior constraints, audit & traceability, exception handling, result verification, and compliance alignment. Each dimension requires concrete engineering mechanisms rather than ad‑hoc prompts.

The proposed solution replaces external fences with an internal semantic skeleton built on an ontology. Three technical pillars support this: architecture constraints that embed permissible actions in the business model, context engineering that injects only the most relevant knowledge sub‑graph into the agent before reasoning, and a feedback loop that validates outputs against the ontology and records traceable evidence.

Knora, the platform demonstrated by 悦点科技, materializes the approach. Its ontology layer stores a label‑property graph (LPG) schema defining five core concepts—Entity, Relation, Event, Action, Logic. The cognition‑engine translates the relevant sub‑graph into the agent’s context and, after execution, compares the result with the ontology to enforce constraints. The execution layer then carries out tool calls and writes results only after passing validation.

A concrete work‑order scenario illustrates the flow: a user requests to change work order WO‑2026‑0312 from 500 to 800 units. The ontology defines a 20 % approval threshold; the change exceeds this, so the engine adds requiresApproval = true, blocks direct write‑back, generates a structured error report, and creates an approval task routed to the designated reviewers. Once approved, the missing approvedBy relation is added, the constraint passes, and the update is committed.

The feedback loop replaces subjective model‑based evaluation with objective rule checking. Hard constraints (e.g., quota limits) are verified deterministically, while soft constraints are delegated to LLM assessment or human review, and every violation feeds back into the ontology to expand coverage and improve future reasoning.

Real‑world deployments in energy transport, electronics manufacturing, finance, and security demonstrate the impact: a rail‑inspection report that previously required 30 people for 7 days is now generated by an agent in 30 minutes with a 3‑person audit, a >70× efficiency boost. Similar gains are reported in quality‑traceability and defect‑analysis pipelines.

Ultimately, the article argues that the competitive moat for enterprise AI will shift from proprietary models to structured business knowledge. Embedding the business ontology creates a self‑evolving semantic map that guides agents, ensures compliance, and continuously improves as agents execute tasks.