Simon Willison, co-creator of the Django web framework and longtime developer blogger, published the opening chapter of a new living guide on March 15, 2026, coining the term "agentic engineering" to describe the practice of building software with the assistance of coding agents. Writing at simonwillison.net, Willison defines agents with deliberate simplicity: "agents run tools in a loop to achieve a goal." He identifies code execution as the single capability that separates agentic engineering from prior LLM-assisted workflows, arguing that the ability to actually run generated code enables iterative verification and convergence on working software. Canonical examples he cites include Claude Code from Anthropic, OpenAI Codex, and Google's Gemini CLI.
Central to Willison's thesis is a reframing of what software engineering actually is. He argues that <a href="/news/2026-03-15-codegen-is-not-productivity-wrong-metric">writing code was never the core of the craft</a> — navigating tradeoffs, specifying problems clearly, and verifying correctness always were. With coding agents absorbing much of the generative work, the human role shifts toward problem specification, <a href="/news/2026-03-14-8-levels-agentic-engineering-framework">tool harness design</a>, and deliberate iteration on instructions. He also flags a key limitation that shapes his practical advice: current LLMs do not learn from past mistakes within a project, so engineers must explicitly encode lessons into updated prompts and tooling rather than expecting the agent to self-correct over time.
The framing carries clear echoes of Willison's earlier work. Django, built in 2003 inside a Kansas newspaper's web team under deadline pressure, was explicitly designed to give small teams leverage through aggressive abstraction and "batteries included" tooling. His subsequent project Datasette applied the same logic to data publishing. Agentic engineering, in his telling, is the next iteration of the same throughline: identify the highest-friction layer in the software development stack and collapse it. The guide's next chapter is titled "Writing code is cheap now" — a phrase that maps directly onto what Django tried to accomplish at the framework level two decades earlier.
The guide, titled "Agentic Engineering Patterns," is explicitly a work in progress. It covers principles, anti-patterns, testing approaches including red/green TDD and agentic manual testing, and techniques for understanding and annotating code. Willison states that no chapter should be considered finished and that he will update content as the field evolves. What sets it apart from the wave of agent tutorials published by tool vendors is the emphasis on limitation rather than capability — particularly the absence of cross-session learning, which Willison treats not as a temporary quirk but as a structural constraint that should shape how the entire discipline is organized.