Machine Teaching for Agentic AI
Most people talk about AI in terms of machine learning. The model is trained on data, patterns are learned, and the system becomes capable of producing useful responses.
That is only part of the story for business AI.
Inside an organisation, the problem is often not that the model lacks general capability. The problem is that the system has not been taught the business. It does not know the client context, the operational standards, the commercial judgement, the governance rules, or the small decisions that experienced people make without writing them down.
This is where machine teaching matters.
Teaching, not just prompting
Prompting asks an AI system to do something now. Machine teaching is different. It transfers expert knowledge into a structured form that the AI can use repeatedly.
In spec-driven AI, the specification becomes the teaching instrument. It captures not only the task, but the intent behind the task:
- the domain knowledge the agent must use
- the rules it must follow
- the judgement criteria it must apply
- the exceptions it must recognise
- the evidence it must provide
- the decisions it must escalate
That specification teaches the system how to behave inside a particular business.
Why this is more valuable than a prompt library
A prompt library can be useful, but it is usually a collection of instructions for isolated moments. It rarely captures the full operating context of a business process.
Machine teaching creates something more durable. It turns senior expertise into a governed asset that can be reviewed, improved, reused, and embedded into workflows. The aim is not simply to get a good answer from the model. The aim is to build a system that keeps producing good work because it has been taught how the business defines good work.
That distinction becomes critical as AI moves from personal productivity into operations.
Teaching agents to work within boundaries
Agentic AI introduces a higher requirement because agents do not only generate text. They can plan, decide, use tools, pass work to other systems, and act across a workflow.
That makes governance inseparable from instruction.
An agent needs to know what outcome to pursue, but it also needs to know the boundaries around that pursuit. Which data can it use? Which decisions can it make? What level of confidence is required? When must a human approve the next step? What evidence must be retained?
Those answers belong in the specification.
From expertise to operating model
The most valuable knowledge in a business often sits in people’s heads: how a partner reads a client situation, how a compliance lead spots risk, how a strategist frames a market, how an operator knows when an exception is no longer an exception.
Spec-driven methodology turns that knowledge into an operating asset. Once captured, validated, and made machine-readable, it can govern agents, workflows, reporting, analysis, and decision support.
The result is not AI that sounds intelligent. It is AI that has been taught to work intelligently inside a specific business.
Read the white paper
This post is adapted from From NASA to AI Agents: The Evolution of Spec-Driven Development. You can also download the PDF.