Mason O’Mara
The Context Transfer Gap
The context transfer gap has two parts. The first is the loss between what exists - a problem, a product, the source - and what you comprehend about what exists. The second is the loss between what you comprehend and what you can clearly communicate. When you delegate work to an AI agent, they operate on what you can communicate, not what exists.
Noise is added at every step in the context transfer gap. Bias and experience in the subject matter obstructs what we are able to comprehend. Our ability and time spent on speaking and writing impacts how we communicate. Agents cannot separate what is noise and what is signal.
Agents add their own noise to each output - needless code, hallucinations, assumptions. The noise compounds, agents operate on previous iteration further away from what exists. The output hardens with each cycle as interdependencies grow and the code becomes dense and more difficult to untangle. The connection between what is being built and signal from what exists is lost.
Eventually you need command over what you built. You learn more about what exists, a new feature is needed, a bug is found. You are blocked by a hardened codebase you either attempt to understand, or cut off and start over. Moving fast and getting feedback is great, moving forward without understanding what was built will fail. How far you can build depends on what you can comprehend and what you can communicate.
How to Work with Agents
Agents help you understand what exists, and what you are building along the way. Agentic code is cheap - you can spin up branches, fast-forward into problems, find solutions, and refactor early. Development time is faster, research and design can be prioritized.
Agents help you move fast when deep understanding isn't necessary - syntax, scaffolding, boilerplate. You can lose signal from what exists quickly, and you can lose understanding of what is being built quickly. Both signal it's time to stop before the code hardens, taking time to rebuild context and understand wahts been build. The code might work, but make it yours before moving forward.
Agents extend waht we build, but they dont repalce understanding the artcitecture, the human intuition, the sngal between gaps that humans feel but cant quantify. That is what our job shifts to focus on. As agents improve and their noise dampens, we prioritize better comprehending and communciating what exists, and moving responsibly with agents.