Two phrases are popping up in conversations about AI: “harness engineering” and “graph engineering”. They sound technical, but the ideas behind them are simple, and they explain why some AI tools are reliable enough for a regulated business while others are not. Here is what each means, in plain English, and why it matters to an RTO.
The model is not the whole system
A large language model on its own is like a very capable new staff member with no handbook, no system access and no supervisor. Useful, but risky. What turns it into something you can trust at work is everything around it: the instructions, the approved documents it can read, the tools it may use, the checks on its work and the points where a human signs off.
What is an AI harness?
That “everything around the model” is what people now call the harness. LangChain put it most simply in March 2026: “If you’re not the model, you’re the harness.” Writing on Martin Fowler’s site, Thoughtworks’ Birgitta Böckeler describes the harness as everything in an AI agent except the model itself.
The phrase took off after an OpenAI engineering post titled “Harness engineering” in February 2026, which described engineers who design environments, specify intent and build feedback loops for AI agents. Anthropic has also written about building effective harnesses for long-running agents. Its November 2025 post notes that agents can declare work finished too early or under-test it, and that structure such as feature lists, progress notes and testing keeps them on track.
Böckeler offers a helpful way to think about it:
- Guides steer the AI before it acts: instructions, approved sources, examples.
- Sensors check its work afterwards: tests, validation rules, a second review.
What does that look like in an RTO?
A few illustrative examples:
- A policy assistant that answers staff questions only from your approved policies and handbook, and shows which document each answer came from. The guide is “only use these documents”; the sensor is “every answer must have a citation”.
- A drafting agent that prepares student emails or feedback, but cannot send anything until a trainer approves it.
- A compliance helper that checks enrolment records against a checklist and flags gaps for a person to review.
This lines up with ASQA’s principles for AI in VET, which say human oversight and accountability must be maintained in all AI-supported activities.
And “graph engineering”?
Be careful with this one. “AI graph engineering” is not yet a standard industry term. A research paper using the label “Graph Engineering” for multi-agent systems only appeared in August 2026. When people use the phrase, they usually mean one of three established ideas:
- Knowledge graphs for AI (GraphRAG). Microsoft Research introduced GraphRAG in 2024. Instead of only searching text for similar passages, it builds a map of the people, things and relationships in your documents. That helps the AI answer questions that join several facts together.
- Workflow graphs. Tools like LangGraph lay out an AI process as steps and decisions, including points where a human can inspect and change what the agent is doing.
- Context graphs. A newer idea, from a December 2025 venture capital essay: a record of the decisions an AI system made and why they were allowed.
Why graphs suit training providers
VET is full of relationships: qualifications contain units, units contain elements and performance criteria, assessments map to criteria, and students produce evidence against them. Modelled as a graph, an AI system could help answer questions like “which assessment covers this performance criterion?” or “what evidence is missing for this student?” This is an illustrative use, not a ready-made product, but it shows where the idea is heading.
Keep expectations realistic
- A graph does not eliminate mistakes. In Microsoft’s own tests, GraphRAG’s faithfulness was similar to standard retrieval, and a 2026 study found it can make results worse in some conditions.
- A harness is only as good as its checks. Human sign-off on anything that affects a student is not optional.
- The field is moving fast. Even experts disagree on how much harnesses will be needed as models improve.
The takeaway
When you assess an AI tool for your organisation, don’t just ask “which model does it use?” Ask what’s around the model. Which sources can it see? What checks its work? Where does a person approve? That’s the harness, and it’s what makes AI safe to use in a regulated business.
Sources
- OpenAI: Harness engineering (Feb 2026)
- Anthropic: Effective harnesses for long-running agents (Nov 2025)
- LangChain: The anatomy of an agent harness (Mar 2026)
- Martin Fowler site: Harness engineering (Apr 2026)
- Microsoft Research: GraphRAG (2024)
- LangGraph overview
- Foundation Capital: Context graphs (Dec 2025)
- arXiv: Graph Engineering in the Era of LLM Agents (Aug 2026)
- arXiv: Towards Robust RAG Based on Knowledge Graph (2026)
- ASQA: Principles for the responsible use of AI in VET