Sebastian Raschka breaks down the “coding agent” — six parts that actually matter

Sebastian Raschka has published a practical primer on what makes modern coding agents work — and it’s not just the model. In a piece titled “Components of a Coding Agent,” he argues that real-world coding performance comes from the system around an LLM: repo context, tools, memory, prompt-cache stability, long-session continuity and model choice. He walks readers through the agent harness concept — the software scaffold that repeatedly calls the model, manages state, and decides what to inspect next — and it has been reported that systems like Claude Code and Codex CLI feel markedly more capable when wrapped in this kind of harness.

The six building blocks, stripped down

Raschka lists six core components that practitioners should focus on: the underlying model (and whether it’s optimized for reasoning), the repository context that grounds the agent in real code, the tool design that exposes code execution and inspection, prompt-cache stability to avoid prompt rot, memory to remember past decisions, and mechanics for long-session continuity. Short answer: size and temperature aren’t enough. You need the rest of the plumbing working in concert.

Why this matters now

So why care? Because the biggest wins in practical developer tooling lately haven’t come from slightly better next-token stats, but from smarter integration: chains of tools, verification steps, and context windows that keep an agent honest. Raschka’s framing — LLM as engine, reasoning model as beefed-up engine, agent as the control loop — makes that point crisp. For developers who’ve ever watched an assistant hallucinate about a file that doesn’t exist, this feels like relief. Finally, not a magic trick but engineering.

Raschka’s article is pitched as a reference for practitioners — a blueprint more than a manifesto. It’s practical, concise, and useful for anyone building or evaluating coding assistants. Allegedly, readers of his books have been asking for exactly this kind of roadmap; now there’s a handy one to point at.

Sources: sebastianraschka.com, Hacker News