Six Key Components of a Coding Agent: How Modern AI Systems for Code Writing Work

The article examines the architecture of coding agents — tools like Claude Code or Codex CLI that serve as a harness around a large language model (LLM). According to the author, in modern systems, it is often the harness, not the model itself, that becomes the key factor determining practical outcomes. He identifies six main components of such a harness. First, live repo context: the agent pre-fetches information about the Git branch, project files, README, and rules (AGENTS.md) to avoid starting from scratch. Second, prompt shape and cache reuse: the stable part of the prompt (general instructions, tool descriptions, workspace summary) is cached and reused, while only the latest user request, recent transcript, and short-term memory are updated. Third, structured tools, validation, and permissions: the agent can only invoke predefined tools (e.g., list files, read file, search, run shell command), which undergo checks for argument correctness, path validity, and the need for user approval before execution. Fourth, context reduction and output management: clipping and history management are used to prevent context growth and degradation of response quality. Fifth, transcripts, memory, and session resumption: interaction logs, notes, and the ability to pause and resume sessions are maintained. Sixth, delegation and bounded subagents: the agent can offload parts of the work to specialized subordinate agents. The author emphasizes that a good harness can significantly boost the effectiveness of even a standard (non-reasoning) model, matching it with more powerful systems if those are used without such a harness.