08-10-2026, 12:10 PM
Components of A Coding Agent
Summary
Sebastian Raschka explains that a coding agent is much more than an LLM: it is a model wrapped in a software “harness” that provides tools, context, memory, and an environment in which the model can repeatedly inspect, reason, act, and verify its work.
He identifies six key components: live repository context, which gives the agent knowledge of files, Git state, and project instructions; prompt structure and cache reuse, which avoids repeatedly processing stable information; structured tools with validation and permissions, allowing safe execution of commands and file operations; context reduction, which clips, deduplicates, and summarizes large histories and tool outputs; session memory, which preserves important information and enables work to continue across sessions; and delegation to bounded subagents for specialized tasks.
The central argument is that much of the apparent intelligence of systems such as Claude Code and Codex comes not only from the underlying model, but from the quality of the surrounding harness—especially its ability to manage context, tools, state, and feedback efficiently.
ARTICLE
Summary
Sebastian Raschka explains that a coding agent is much more than an LLM: it is a model wrapped in a software “harness” that provides tools, context, memory, and an environment in which the model can repeatedly inspect, reason, act, and verify its work.
He identifies six key components: live repository context, which gives the agent knowledge of files, Git state, and project instructions; prompt structure and cache reuse, which avoids repeatedly processing stable information; structured tools with validation and permissions, allowing safe execution of commands and file operations; context reduction, which clips, deduplicates, and summarizes large histories and tool outputs; session memory, which preserves important information and enables work to continue across sessions; and delegation to bounded subagents for specialized tasks.
The central argument is that much of the apparent intelligence of systems such as Claude Code and Codex comes not only from the underlying model, but from the quality of the surrounding harness—especially its ability to manage context, tools, state, and feedback efficiently.
ARTICLE
┌────────────────────────────────┐
│ KONSTANTINOS MICHAILIDIS │
└────────────────────────────────┘
│ KONSTANTINOS MICHAILIDIS │
└────────────────────────────────┘