Version 3

SemiAnalysis argued in July 2026 that agentic coding harnesses — Claude Code, Codex, OpenCode — are context orchestration tools sharing a common 'plan, act, verify' loop, and that model quality is the decisive variable in agentic performance. [3][2][1] Practitioners pushed back, citing harness-layer failures as the dominant operational constraint in real deployments. [7][8] Simon Willison, amplifying Armin Ronacher, reported that newer Anthropic models Opus 4.8 and Sonnet 5 perform worse on custom edit tool schemas than their predecessors — theorized as RL training optimizing for Claude Code's native built-in tool at the cost of third-party schema conformance. [10]

The Willison/Ronacher finding moves the model-vs-harness debate from first principles to a concrete product problem: if RL training couples models to the primary harness's native tool schemas, third-party harness builders may find that capability improvements come at a cost to their specific implementations. Where investment goes — model quality or harness engineering — depends on which layer is the binding constraint, and the answer may differ by harness and use case.

On July 3, 2026, SemiAnalysis published a thread aimed at technically-minded practitioners asking what agentic coding harnesses actually are. [1] The core argument: LLMs like Opus and GPT-5.5 are stateless — they retain no memory between requests, and the harness reconstructs and resends the full conversation on every turn. Every request contains the same three parts: a system prompt, tool definitions as JSON schemas, and a chronological message history. [2] What distinguishes Claude Code from Codex or OpenCode is therefore differing choices about context management strategy, not some distinct underlying architecture. The thread's conclusion: harnesses are often overhyped, and model quality is the real determinant of agentic performance — 'everything else is just REST all the way down.' [3]

Tool use, as SemiAnalysis describes it, is a simple loop: when the model decides to invoke a tool, it generates a JSON object specifying the call; the harness parses that JSON, executes the tool, and returns the output in the next turn. [4] This repeats through a 'plan, act, verify' cycle that SemiAnalysis argues all harnesses share. Model-routing tools have emerged alongside this debate — Rayline routes Claude Code subagents to cheaper models [5], and Workweave handles routing within Claude, Codex, and Cursor [6] — suggesting that decoupling harness from model selection is itself a recognized engineering problem.

That conclusion meets direct pushback from practitioners. Prasenjit Sarkar argues the bottleneck in agentic coding has moved from the model to the harness layer. [7] Duy (goon_nguyen) adds from production experience that when agents break for real users, the model is usually not the first failure point — retry logic, context limits, tool routing, and environment handling fail first. [8] AiDevCraft observes that model-swapping under a fixed harness still yields usable agentic coding on consumer hardware, implicitly supporting the model-primacy view. [9]

The sharpest evidence against clean model-harness separability comes from Simon Willison, amplifying Armin Ronacher's finding: newer Anthropic models Opus 4.8 and Sonnet 5 are worse than their predecessors at conforming to custom edit tool schemas — they invent extra fields not present in the schema. [10] The proposed explanation is that RL training optimized these models for Claude Code's built-in search-and-replace edit tool, which then interferes with differently shaped schemas used by third-party harnesses. OpenAI's Codex, which uses a distinct edit mechanism (apply_patch) and has been trained on it, illustrates the same dynamic from the other direction. For third-party harness builders, the implication is that 'better' models may perform worse on their specific tool implementations — concrete evidence that the model and harness layers are more tightly coupled than the 'REST all the way down' framing implies.