At a recent AIE event, Geoffrey Litt framed a practical challenge for people collaborating with coding agents: you need to understand the code deeply enough that you can continue to participate in the project. Litt summed this up as "Understand to participate," warning that as agents make increasingly large and sophisticated changes, people's mental models can drift away from how the code actually works, creating what he calls cognitive debt.
The risk of cognitive debt
Litt cautioned that when understanding diverges from the code's real behavior, teams incur cognitive debt — gaps in knowledge that later limit the ability to shape and maintain the project. Avoiding that debt requires more than operating the agent; it requires learning what the agent is doing well enough to be an active creative partner.
Why conceptual fluency matters
According to Litt, a rich set of concepts in your mind enables creative and fluent thinking about how to move a project forward. Without that fluency, a person's ability to participate meaningfully is significantly constrained:
- you must be able to see what the agent is doing to contribute to the creative process;
- a well-developed mental model helps you decide next steps and design changes;
- lacking an internal model reduces decision-making and planning capability.
Practical implications
From Litt's remarks, the practical implication for developers and product teams is clear: explicitly learn and internalize the code and the agent's behavior, even if much of the code is produced by automated agents. Doing so not only helps prevent errors but also increases the effectiveness of collaboration and preserves creative control.
Availability and further reading
The AIE talks are all recorded — more than 300 sessions — and the organizers plan to release them over the next three weeks. Geoffrey Litt's talk is recommended viewing on YouTube, and he also published a thread version of his talk on Twitter.
The issue Litt raised is central to the practice of generative AI and large language models (LLMs): the quality of human–AI collaboration depends heavily on how well humans understand the underlying code and decision logic.
