Simon Willison (co-creator of Django and software developer) breaks down the architecture of coding agents, revealing them as sophisticated software harnesses for Large Language Models. At their core, these agents interpret text as numerical tokens and use chat-templated prompts to simulate ongoing conversations. Because models are inherently stateless—meaning they forget previous interactions—the harness must replay the entire conversation history for every new request.

To make these models useful for programming, developers equip them with "tools," which are essentially functions the model can trigger via specific text patterns. This allows an agent to execute terminal commands or run Python scripts to verify its work. Efficiency is maintained through token caching, a technique that reduces costs by reusing calculations for common segments of input data that have been processed recently.

The most significant evolution in coding agents is the introduction of "reasoning" phases. Instead of generating code instantly, the model spends extra tokens to "think" through a problem, weighing different solutions before providing a final answer. This iterative process mimics human problem-solving and is particularly effective for navigating deep codebases and complex debugging scenarios where first-guess solutions often fail to account for edge cases.

Simon Willison (co-creator of Django and software developer) breaks down the architecture of coding agents, revealing them as sophisticated software harnesses for Large Language Models. At their core, these agents interpret text as numerical tokens and use chat-templated prompts to simulate ongoing conversations. Because models are inherently stateless—meaning they forget previous interactions—the harness must replay the entire conversation history for every new request.

To make these models useful for programming, developers equip them with "tools," which are essentially functions the model can trigger via specific text patterns. This allows an agent to execute terminal commands or run Python scripts to verify its work. Efficiency is maintained through token caching, a technique that reduces costs by reusing calculations for common segments of input data that have been processed recently.

The most significant evolution in coding agents is the introduction of "reasoning" phases. Instead of generating code instantly, the model spends extra tokens to "think" through a problem, weighing different solutions before providing a final answer. This iterative process mimics human problem-solving and is particularly effective for navigating deep codebases and complex debugging scenarios where first-guess solutions often fail to account for edge cases.