SemiAnalysis Demystifies Agentic Coding Harness Architecture: Model vs. Orchestration

On July 3, 2026, SemiAnalysis published a thread demystifying agentic coding harnesses — Claude Code, Codex, OpenCode, and similar tools — arguing they are fundamentally context orchestration tools, not architecturally distinct systems. [1] Every harness request to a stateless LLM contains the same three components: a system prompt, tool definitions as JSON schemas, and a chronological message history; what differs between products is context management strategy, not underlying structure. [3][2] SemiAnalysis concludes that model quality, not harness engineering, is the decisive variable in agentic performance. [5] Practitioners push back directly, arguing the operational bottleneck has in practice moved to the harness layer. [6][7]

How the industry frames the model-vs-harness split determines where engineering investment and vendor differentiation actually go. If harnesses are commodity wrappers, model quality is the competitive variable; if harness design is the primary lever, orchestration tooling and context management become the field worth owning.

On July 3, 2026, SemiAnalysis published a thread aimed at a technically-minded audience asking what agentic coding harnesses actually are. [1] The thread's foundational premise is that major LLMs — Opus, GPT-5.5, and others — are stateless: they retain no memory between requests, and the harness must reconstruct and resend the full conversation on every user turn. Prompt caching exists specifically to reduce the computational cost of this repeated context assembly. [2] Every request the harness builds contains the same three parts: a system prompt, tool definitions expressed as JSON schemas, and a chronological message history. What distinguishes Claude Code from Codex from OpenCode is therefore not some distinct architecture but differing choices about how that context is managed. [3]

The mechanics of tool use, as SemiAnalysis describes them, are a simple loop: when the model decides to invoke a tool, it generates a JSON object in its response body specifying the call. The harness parses that JSON, executes the tool on the host machine, and returns the output to the model in the next turn. [4] This repeats through a 'plan, act, verify' cycle that SemiAnalysis argues all harnesses share. The thread's conclusion is pointed: harnesses are often overhyped, and the real power in agentic systems is still in the underlying models — 'everything else is just REST all the way down.' [5]

That conclusion meets direct pushback from practitioners. Prasenjit Sarkar argues the bottleneck in agentic coding has moved from the model to the harness layer. [6] Duy (goon_nguyen) makes a related point from production experience: 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. [7] AiDevCraft, by contrast, observes that swapping the underlying model under the same harness shape still yields usable agentic coding on consumer hardware (a 3090), which implicitly supports the SemiAnalysis framing by suggesting the harness shape itself is not the binding constraint. [8]

The surrounding ecosystem adds texture. Model-routing tools have emerged specifically to decouple harness from model selection: Rayline routes Claude Code subagents to on-device and cheaper models [9], and the Workweave router handles routing within Claude, Codex, and Cursor. [10] A reported leak of Claude Code's source code, followed by an overnight community rewrite [11], drew attention to how much actual complexity may live in the harness — a data point that sits uneasily with the 'REST all the way down' framing. Separately, Harness (the DevOps company, distinct from the architectural concept) shipped autonomous worker agents on June 30 [12], illustrating that the commercial market is already building on these primitives regardless of where the theoretical debate settles.