Ontology-Driven Agent Control: Semantic Foundations for Harness Engineering

From Agent Hype to "Uncontrollable"

In 2024-2025, Agents became the primary form of enterprise AI deployment. They can autonomously plan, call tools, and execute multi-step tasks, performing well in demos. However, in real business they exhibit similar problems: inaccurate terminology, deviating reasoning logic, execution results misaligned with enterprise rules, and "confidently doing the wrong thing" at critical nodes.

The root cause is not insufficient model strength, but that Agents lack a "structure that understands the rules" — they know how to work but not the industry boundaries or enterprise decision rules. This context drove the rise of "Harness Engineering" in Q1 2026, whose core is to establish a complete constraint, feedback, and control system for Agents so they can act autonomously without crossing business boundaries. The industry has converged on three technical pillars: architectural constraints to limit behavioral boundaries, context engineering to manage information input, and feedback loops to verify output quality.

Most practices remain at the engineering assembly level — writing prompts, listing rules, building workflows, configuring permissions. These work in simple scenarios but expose a deep issue in complex business: constraints are "externally added", semantics are "implicit", and compliance relies on the model's understanding of rules, which is itself unreliable.

This article explores a different path: using Ontology to provide semantic infrastructure for the Harness system, making constraints derive from business structure itself rather than engineering configuration. Ontology explicitly models business semantics, providing a unified structured foundation for the three pillars.

Redefining the Problem: "Safe and Controllable" as a Multi-Dimensional Engineering Proposition

Before detailing the solution, the article clarifies the boundaries of "safe and controllable execution" across six independent yet interrelated dimensions:

Permissions & Isolation — Who can do what? Can data cross boundaries? Engineering means: RBAC/ABAC, API gateway, data sandbox.

Behavioral Constraints — Where are the Agent's reasoning and invocation boundaries? Engineering means: Prompt constraints, tool whitelist, ontology modeling.

Audit & Traceability — What was done? Can the decision process be reconstructed? Engineering means: Operation logs, decision chain tracing, explainability frameworks.

Exception Handling — How to degrade and roll back when errors occur? Engineering means: Circuit breakers, human review nodes, idempotent design.

Result Verification — Does output comply with business rules? Engineering means: Rule engines, formal verification, ontology constraint validation.

Compliance Alignment — Does it meet industry regulatory requirements? Engineering means: Compliance knowledge bases, approval flow integration, auditable reports.

The ontology-driven solution primarily acts on "Behavioral Constraints" and "Result Verification", with substantive contribution to "Audit & Traceability". It serves as the semantic infrastructure layer within a complete safety system, not a replacement for other engineering measures.

Architectural Constraints: From "External Fences" to "Internal Skeleton"

Engineering constraint methods face three structural difficulties in complex business: rule count grows with complexity, inflating maintenance costs; rules expressed in natural language introduce model understanding uncertainty and bypass risk; semantic relationships between rules and business objects are implicit, preventing cross-scenario reuse.

Ontology differs by not trying to "block" the Agent but defining its behavioral space from the start. Constraints become an internal skeleton built into the business structure, not an external fence.

Key mechanism: business rules are no longer natural language in prompts but explicitly modeled as queryable, verifiable structures. Tools are managed uniformly at the ontology layer — what the Agent can use, how to trigger, and execution processes are dictated by structure, not the model. Validation occurs after Agent output but before operation execution: the system compares the Agent's execution intent against the ontology; if rules are violated, it rejects and forces re-reasoning. Validation bases are specific traceable nodes and relationship paths in the structure, yielding deterministic conclusions independent of model understanding.

Context Engineering: From "Supplementing Memory" to "Restructuring Memory"

Agents frequently "lose memory" in long tasks — repeatedly asking for basic information or losing context at critical nodes. The root cause is linear text stacking without structure. Adding external memory, context compression, or extending window length only alleviates, not solves, the problem.

Ontology's entry point: enterprise data, processes, and relationships are inherently a complex association network, not linear. Ontology makes these relationships explicit, building a queryable, evolvable business semantic network.

This structure brings three substantive improvements:

Precise Retrieval Replaces Full Injection : Before Agent startup, the cognitive engine extracts a task-relevant semantic subgraph from the ontology, dynamically injecting only the most relevant context into the reasoning context, controlling context overflow and irrelevant interference at the source.

Consistency Assurance : Business knowledge is managed by a unified semantic network; expired information, conflicts, and redundancies are systematically handled. The Agent always reasons on the latest consistent information, not potentially outdated prompt fragments.

Cross-Task Reuse : The same semantic structure serves different task Agents, eliminating per-scenario context reorganization. The ontology is a continuously maintained "business map" on which Agents act, rather than starting from zero each task.

Additionally, ontology-driven context engineering resolves the long-standing split between symbolic reasoning and LLM reasoning. Traditional knowledge graph reasoning is explainable but limited in coverage; pure LLM reasoning is flexible but uncontrollable. Ontology provides a collaboration framework: where structure covers, ontology gives deterministic constraints; where structure doesn't cover, LLM can supplement but conclusions are explicitly marked with confidence status. Both play their roles, preserving both accuracy and flexibility.

Feedback Loop: From "Subjective Evaluation" to "Traceable Verification"

Mainstream feedback mechanisms introduce an "evaluator" model to judge another model's output. This has some effect but clear limitations: models are easily persuaded by superficially reasonable results and lack judgment on business-level correctness.

Ontology provides a more direct path: objective verification based on structure, not subjective evaluation. Many enterprise judgments are essentially rule-based — whether exceeding quotas, satisfying preconditions, or complying with process logic. These don't require "understanding" and can be directly verified. Once rules are encoded in ontology, every Agent output can be automatically compared against business constraints, yielding a binary pass/fail.

Not all judgments can be formalized. "Hard constraints" suit ontology verification; "soft constraints" do not — e.g., whether a decision follows business convention, whether abnormal motives exist behind an operation, or the definition of "major change" itself involves subjectivity. For soft constraints, the ontology approach combines with LLM evaluation or human review nodes, complementary rather than mutually exclusive.

Another value of the feedback loop is continuous ontology evolution. The relationship between ontology and Agent is not unidirectional constraint: ontology defines Agent capability boundaries; Agent encounters massive business data in real execution, which feeds back to ontology — identifying uncovered concepts and relationships, marking frequent error locations in reasoning paths, guiding ontology correction and iteration. This continuous evolution loop is the core difference from traditional "build and forget" knowledge graphs.

Explainability is another critical value. Every verification conclusion has explicit structural basis; every judgment traces to specific rule nodes. Regulatory audits need "system can prove compliance", not "model thinks it's compliant". This gap is decisive in strong-compliance industries like finance, industrial, and railway.

From Technical Control to Business Control: Knora's Implementation Path

1. Constraints Originate from Business Itself, Not Engineering Assembly

Traditional Harness solutions make Agents "more stable", but stability boundaries are engineering boundaries — rules written by humans, constraints configured, compliance maintained via prompts. The ceiling is clear: once business complexity exceeds engineering maintenance capacity, constraints fail.

Ontology-driven approach achieves something else: constraints come directly from business structure. Enterprise objects, relationships, and rules are uniformly modeled as ontology. When executing tasks, the Agent doesn't just receive context and generate; it must first enter this semantic layer, reasoning and deciding within predefined business relationships. It operates on a defined "business map", not free-form.

This shift brings three substantive transformations: Agents become business-semantic participants, not just execution tools; the system reuses constraints on a unified semantic foundation instead of rebuilding per scenario; business rule layer remains stable during model iterations, avoiding drift with model versions. This is the leap from "technical control" to "business control".

2. Knora's System Architecture

YueDian Technology's Knora platform implements the methodology as a layered collaborative architecture.

Ontology Layer (Knowledge Foundation) : Base layer stores ontology model as a Labeled Property Graph (LPG). The meta-schema defines five core concepts: Entity (business objects like work orders, equipment, personnel), Relation (semantic connections between entities), Event (meaningful state changes in business), Action (system-executable operations with trigger conditions and parameter constraints), and Logic (execution flows defined by a DAG orchestration engine).

Key design: traditional ontologies mainly solve "what is" classification; Action and Logic turn ontology from static knowledge description into executable business specifications that directly drive Agent tool calls and process orchestration. Tools are defined at the ontology layer, not the Agent layer — what tools the Agent can use, how to trigger, and execution processes are all decided by ontology. This layer is the system's knowledge foundation, continuously evolvable.

Cognitive Engine Layer (Translation & Arbitration) : Bridges ontology and Agent. Two core roles: (1) Before Agent startup, reads domain knowledge relevant to the current scenario from ontology — entity relationships, business rules, available tools — and injects into Agent's reasoning context, so the Agent knows both "how this domain operates" and "what tools I can use and how". (2) After Agent generates results, brings conclusions back for ontology constraint validation; if ontology-defined rules are violated, rejects and forces re-reasoning.

Agent Execution Layer (Task Executor) : Autonomous intelligent agent product layer. Receives user tasks, calls tools, generates results, completes execution. This layer is a pure executor — tool sources, usage, boundaries are not self-decided but defined by ontology layer and passed via cognitive engine.

Data flow: User task → Cognitive engine extracts relevant knowledge from ontology and injects into context → Agent reasons and executes in this context → Results return to cognitive engine for ontology validation → Output only after passing. Ontology does not directly interact with Agent; it acts through the cognitive engine intermediary.

Manufacturing example: "Work order quantity change requires approval". Full execution chain:

(1) User intent: "Change work order WO-2026-0312 planned quantity from 500 to 800".

(2) Cognitive engine queries ontology (LPG graph traversal): Node WO-2026-0312, type WorkOrder, status "Issued/Not Started"; attribute change magnitude 60%, exceeding ontology-defined 20% approval threshold; Action rule: over-threshold change triggers condition requiresApproval = true; Relationship chain:

WorkOrderChange → approvedBy → [BOMEngineer, ProductionManager]

.

(3) Constraint validation: Current operation lacks approvedBy relationship, validation fails.

(4) System response: Block direct write; generate structured error report with violated rule nodes and relationship paths; auto-create approval task routed to corresponding approvers; write operation intent to audit log with status "Pending Approval".

(5) After approval: Add approvedBy relationship, re-validate pass, execute write.

3. Automated Modeling: Layered Processing and Human-Machine Collaboration

Enterprise ontology cold-start cost is a real challenge. Manual construction of sufficiently covering ontology traditionally takes weeks or months, requiring professional knowledge engineers.

Knora's strategy is not "fully automatic in one step" but layered processing, confidence-driven, human-machine collaboration fallback. Structured, regular tasks (e.g., field-to-ontology-attribute semantic mapping, auto-generating data governance DAG flows from ontology metadata) are automated. Tasks involving business semantic judgment and concept boundary definition are not forced automated. System assigns confidence scores to each auto-generated result: high confidence executes directly, medium confidence recommends human confirmation, low confidence enters review queue. Human intervention targets only truly ambiguous parts. Each human confirmation or correction feeds back to the system, continuously improving subsequent auto-modeling capability. As industry templates accumulate and domain samples adapt, human intervention ratio gradually decreases.

4. Real-World Deployments

Knora has deployments in energy transportation, electronic manufacturing, finance, and security sectors.

Energy Transportation : Railway comprehensive inspection report generation Agent automatically constructs indicator systems and generates reports. Work that originally required 30 people / 7 days to organize data and write reports becomes 3 people spending one day verifying data, Agent auto-completes in 30 minutes — efficiency improvement over 70x. In safety hazard identification, maintenance plan generation, emergency plan generation, emergency resource scheduling, and dozens of other scenarios, it serves as "digital employees" undertaking actual business responsibilities.

Electronic Manufacturing : Around core scenarios like quality traceability and defect analysis, transforms experience-dependent manual processes into precise knowledge-driven automated digital business flows.

Conclusion

AI's entry into enterprises is undergoing a fundamental fork. One path continues stacking tools, accumulating prompts, doing integration — Agents run but no one truly knows where they'll go or what results they'll produce. The other path builds business structure first, letting Agents act on a clearly defined semantic map — they know where boundaries are, what rules are, and the basis for every decision step.

The short-term cost difference between the two paths is small, but the gap after one year, three years will manifest in every moment needing to explain decisions to regulators, every moment of semantic confusion in multi-system collaboration, every moment an Agent "confidently does the wrong thing" at a critical node.

Enterprise AI competition is ultimately not between models, but who structures their business knowledge first . Enterprises that close this loop first accumulate not just a system, but a continuously self-evolving business intelligence foundation — becoming more accurate with every Agent execution, more complete with every business iteration.

The real moat is never bought tools, but precipitated cognition. Agents will be replaced, models will be iterated, but business cognition沉淀 in ontology will not — that is the reason to start now.