IDD concept: Govern execution
The Human Layer and the AI Layer
Intent-Driven Development separates delivery into two layers. In the Human Layer, people define intent, set governance and carry accountability. In the AI Layer, autonomous systems carry out implementation and automation. Keeping the two apart is what allows the methodology to survive the rapid evolution of AI models and tools.
The coupling problem
Many approaches to AI in delivery fail for the same reason: they couple human concerns directly to the capabilities of current AI. Prompt engineering techniques that were carefully learned became obsolete within a couple of years, and architectures built around a particular tool needed reworking every time the market shifted. When a framework is coupled to the technology, every significant change in the technology forces a restart.
The separation principle
IDD enforces a clear separation between the two layers, with a governance interface between them.
- The Human Layer: stable
- User needs, domain models, Intent Specifications, stakeholder communication and accountability. These change when the business changes, not when a new model is released.
- The governance interface
- Risk Dials, Human Gates and Intent Fidelity measurement, which together decide how much autonomy each part of the AI Layer receives.
- The AI Layer: fluid
- Implementation tools, model capabilities, agent architectures and levels of automation, all of which can change quickly and frequently.
When AI improves, the organisation adjusts the governance controls. When tools change, it swaps the implementation. The Human Layer stays stable, the AI Layer remains fluid, and the interface absorbs the change.
Why specifications survive change
Because Intent Specifications live in the Human Layer, they describe what to build and why, not how the AI should organise itself. They are tool-agnostic, model-agnostic, version-agnostic and architecture-agnostic. Moving from a single agent to a multi-agent system, from one vendor to another, or from today's models to whatever follows them changes the AI Layer and leaves the specification untouched.
The same separation gives organisations a safe way to adopt new capabilities. When something new emerges, governance controls stay where they are while the new implementation is evaluated in parallel; fidelity is measured rather than assumed; and improvements are introduced first in low-risk areas. Learning compounds instead of restarting with every shift in technology.
Build on bedrock. Let the AI Layer evolve. Your investment compounds.
Humans have roles, agents have functions
The two layers are organised differently. A human role carries accountability, decision-making authority, domain knowledge and judgement. An agent performs a function, something narrow such as generating code, running tests, deploying a service or monitoring a system. Agents do not need job titles, career paths or reporting lines, and the AI Layer is most naturally designed around functions rather than as an imitation of a human organisation chart.
That distinction also guards against a common trap: inserting agents into existing structures and relabelling the roles around them, so that a business analyst becomes an "intent analyst" and a developer becomes a "prompt engineer". That keeps the old organisation with new names. In IDD, people move up a level instead, from how to what and why: setting direction, defining outcomes and constraints, reviewing results, making trade-offs and governing behaviour.
Where it comes from
The separation principle was set out in Why IDD Survives Rapid AI Model Evolution on 2 February 2026, applied to multi-agent systems in IDD via Multi-Agent Systems, and completed by Humans Have Roles, Agents Have Functions. It forms Chapters 3, 7 and 14 of the ebook.