Coding agents and their underlying models are becoming more capable, with larger context windows. These advances should allow us to get out of the agent’s way and let it build independently. But that only works if what it builds is what we intended. Larger context windows do not make structured context less important. If anything, poorly articulated intent can allow an increasingly capable agent to travel further in the wrong direction before we intervene. The result is more clarification, review, correction, and potentially expensive rework. The bottleneck is our ability to articulate intent clearly and provide the right context so that agent-generated implementations stay aligned with our goals.

Every time we leave something unstated, ambiguous, or disconnected from the context available to an agent, we leave room for interpretation. Intent-Driven Development is about engineering this interpretation gap towards zero. In this talk, we will build a harness around this principle. We will use techniques such as Spec-Driven Development and iterative loops to articulate and refine intent. We will structure context so that agents can navigate relationships between requirements, decisions, terminology, constraints, and implementation. We will use artifacts such as ADRs and glossaries to preserve important context and make the necessary information available when the agent needs it, rather than putting everything into the context window upfront. Most importantly, we will use feedback from agent execution to detect where interpretation entered the system and continuously improve the harness to prevent the same deviations from recurring.

The goal is to move towards a software factory where we do not have to constantly supervise agents to keep them on track. We engineer and continuously tune the system so that the software produced by the harness converges towards what we intended. The less we leave open to interpretation, the more autonomy we can safely give the agent.