Six Key Components of a Coding Agent: How Modern AI Systems Write Code

In a detailed article, Sebastian Raschka explains the design of coding agents and harnesses, using his Mini Coding Agent as a concrete example. He distinguishes between LLMs, reasoning models, agents, and agent harnesses. The six core components are: 1) live repo context (collecting workspace facts like git branch, README, etc.), 2) prompt shape and cache reuse (separating stable prefix from dynamic parts to avoid reprocessing), 3) structured tools with validation and permissions (known tools with validated arguments and safety checks), 4) context reduction and output management (clipping irrelevant history to control bloat), 5) transcripts, memory, and resumption (session store, short-term memory, ability to resume tasks), and 6) delegation and bounded subagents (tools that can delegate work to sub-agents). The author also notes that the harness can be the distinguishing factor between models, as modern LLMs have similar raw capabilities.