← Articles

Co-designing with coding agents

Studio case note · June 2026 · Updated September 2026

I run Good Design as one person. The studio makes videos, websites, and apps, and carries more projects than I could coordinate alone. Agentic workflows help carry the context, coordination, and repeatable execution, leaving me more attention for creative direction and judgment.

Co-design gives the human and the agents different responsibilities. I set the intent, choose what matters, and judge the result. Coding agents help turn those decisions into working software; other agents support research, production, and organization. The useful distinction is between the decisions that need my authorship and the work a system can carry through with clear direction.

I describe what I want a thing to be. An agent proposes a structure, drafts a specification, and identifies the decisions that need my judgment. I choose. The approved decisions enter the shared project record so the next session can build on them. The work moves through a loop of intent, execution, review, and revision.

Co-design also depends on feedback. Donald Schön described the situation talking back to the designer: the material answers, resists, and changes the next move (Schön, 1983). In an agentic workflow, that feedback can be a working prototype, a failed test, or a conflict with an earlier decision. The agent should make the conflict visible before proceeding. Edwin Hutchins called thinking spread across people, tools, and artifacts distributed cognition (Hutchins, 1995); the designer, agents, and shared project record can be understood as a joint cognitive system (Hollnagel & Woods, 2005).

My attention is the scarce resource. Herbert Simon observed that a wealth of information creates a poverty of attention, so the workflow has to spend it carefully (Simon, 1971). Routine work should proceed within an agreed scope. Decisions that need me should arrive with enough context to judge them. The measure is how much meaningful work I can direct and evaluate, not how many tasks an agent can generate.