The Coding Agent Is Not the Model — It's the Harness编程智能体真正的核心：不是模型，是 Harness

Frontier base models have converged in raw capability; the gap between a great coding tool and a mediocre one now lives almost entirely in the harness — the software layer that manages context, tools, caching and memory around the model.

Raschka decomposes a coding harness into six components: live repo context, cacheable prompt structure, validated tool calls, context-bloat control, two-tier session memory, and bounded subagent delegation.

Raschka opens by separating three overloaded terms people conflate when discussing 'LLM coding ability': the LLM itself (next-token predictor), the reasoning model (an LLM post-trained or prompted to spend extra test-time compute on intermediate steps and self-verification), and the agent (a control loop wrapped around either). His car analogy — LLM as engine, reasoning model as a tuned but pricier engine, harness as the vehicle that lets you actually drive it — sets up the article's real thesis: the harness, not engine tuning, is where most of the perceived capability gain now lives.

He argues frontier base models (GPT-5.4, Opus 4.6, GLM-5-base) have essentially converged in raw capability, so the same model dropped into a strong harness would likely perform close to Claude Code or Codex — though harness-specific post-training (OpenAI's separate GPT-5.3 vs. GPT-5.3-Codex) still adds marginal gains. This reframes 'which model is smarter' debates as largely secondary to 'whose harness is better engineered.'

The six components map onto a coherent pipeline rather than isolated features. Repo context (1) and prompt caching (2) are both about giving the model grounded, cheap-to-reuse situational awareness before it acts — separating 'what facts to gather' from 'how to package and cache them.' Tool access (3) is the layer that converts model output into real, gated actions, with schema validation, path restriction, and human-approval gates acting as a security perimeter around an inherently unpredictable text generator.

Context management (4) and session memory (5) are subtly different despite both dealing with 'history': slimming is about what to feed back into the next prompt (clip long outputs, dedup repeated file reads, summarize aggressively with recency bias), while structured memory is about what to persist to disk long-term — a full transcript for lossless recovery plus a small, actively curated working-memory layer for task continuity. Conflating these two would either bloat every prompt or lose recoverability.

Subagent delegation (6) is presented as the natural next step once tools and memory exist, but Raschka is careful to flag its main failure mode: unconstrained subagents duplicate work, race on the same files, or spawn further subagents recursively. The fix — inheriting just enough context while imposing strict boundaries (read-only access, capped recursion depth) — is where he contrasts Claude Code's stricter read-only default against Codex's looser approach of inheriting the parent's sandbox and approval permissions, scoping subagents instead by task, context size, and depth.

Compared to OpenClaw, a general-purpose local agent platform, coding harnesses are narrowly optimized for one job — inspecting and editing a codebase and running local tools — whereas OpenClaw runs many long-lived agents across chats and workspaces, treating coding as just one of many chores. This closing contrast sharpens the article's definition: a coding harness is a special case of agent harness, purpose-built around software engineering's specific information needs.