The Harness Effect: How Orchestration Design Sets
the Token Economics of Enterprise Agentic AIHarness 效应：编排设计如何决定企业智能体 AI 的 Token 经济学

Let a task execute as a $k$-turn agent loop. Turn $i$ submits input tokens $T^{\mathrm{in}}_{i}$ and receives output tokens $T^{\mathrm{out}}_{i}$. With input price $p_{\mathrm{in}}$ and output price $p_{\mathrm{out}}$ (per token), the task costs假设任务执行 k 轮 Agent 循环。第 i 轮提交输入 Token Tiin，接收输出 Token Tiout。输入价格 pinp，输出价格 pout，任务成本为...

$$C\;=\;\sum_{i=1}^{k}\Big(p_{\mathrm{in}}\,T^{\mathrm{in}}_{i}+p_{\mathrm{out}}\,T^{\mathrm{out}}_{i}\Big).$$

(1)

The input side decomposes into terms the harness constructs:输入侧可拆解为 Harness 构建的项：

$$T^{\mathrm{in}}_{i}\;=\;\underbrace{S_{i}}_{\text{system}}+\underbrace{H_{i}}_{\text{history}}+\underbrace{G_{i}}_{\text{tool schemas}}+\underbrace{R_{i}}_{\text{retrieval}}+\underbrace{U_{i}}_{\text{user turn}},$$

(2)

and the loop multiplies everything by retries and dead-end branches, which the harness also governs. A naive harness replays the full transcript: $H_{i}=\sum_{j<i}(U_{j}+T^{\mathrm{out}}_{j}+O_{j})$ where $O_{j}$ is tool output at turn $j$. Total input tokens then grow quadratically in turn count,循环将一切乘以重试和死胡同分支，这也是 Harness 管辖的。笨拙的 Harness 重放完整记录：Hi=∑j<i(Uj+Tjout+Oj)，其中 Oj 是第 j 轮的工具输出。总输入 Token 随轮次平方级增长。

$$\sum_{i=1}^{k}T^{\mathrm{in}}_{i}\;\approx\;kS\;+\;\frac{k(k-1)}{2}\,\bar{m}\;+\;k\bar{G}\;+\;\textstyle\sum_{i}R_{i},$$

(3)

with $\bar{m}$ the mean per-turn payload. A harness that compacts history, caches the invariant prefix (tool schemas included), offloads bulky tool outputs, and truncates retrieval to the evidential minimum converts the quadratic term to (approximately) linear (Figure 1). Nothing about the model changes; the bill does.m¯ 是平均每轮负载。如果 Harness 压缩历史、缓存不变前缀（含工具定义）、卸载大块工具输出、将检索截断至证据最小值，就能将平方级增长转为线性（图 1）。模型没变，账单变了。

Two further facts sharpen Eq. (1). First, agent workloads are input-dominated: because the transcript is re-submitted on every turn, production agents report input-to-output token ratios near 100:1 [30], so the $p_{\mathrm{in}}$ term is nearly the whole bill. Second, the price of an input token is not one number. Providers serve tokens that repeat a previously seen prompt prefix from cache at a deep discount—roughly $0.1\times$ the base input rate [32, 30]. If a fraction $h$ of input tokens are cache reads billed at multiplier $\kappa$, the effective input price is两个事实加剧了 Eq.(1)。首先，Agent 负载是输入主导的：因为记录每轮重发，生产 Agent 的输入输出比接近 100:1，所以 pinp 项几乎就是全部账单。其次，输入 Token 价格不统一。提供商对缓存命中部分有深折扣——约为基础输入价的 0.1 倍。如果比例 h 的输入 Token 是缓存读取，有效输入价格为...

$$p^{\mathrm{eff}}_{\mathrm{in}}\;=\;p_{\mathrm{in}}\big(1-h\,(1-\kappa)\big),\qquad\kappa\approx 0.1,$$

(4)

so a harness that holds $h$ near $1$ pays roughly a tenth of list price for the dominant term of Eq. (1). Crucially, $h$ is neither a model property nor a provider favor: it is a function of prompt byte-stability across turns, which is set entirely by how the orchestration layer assembles context. The harness therefore controls both factors of the bill—how many tokens are submitted (Eq. (2)) and the price at which the dominant ones are billed (Eq. (4)).所以，将 h 保持在 1 附近的 Harness，其核心项成本仅为定价的十分之一。关键在于，h 既不是模型属性，也不是提供商恩赐，而是由编排层如何组装上下文决定的。Harness 因此控制了账单的两个因子——提交了多少 Token，以及核心项按什么价格结算。

Define token intensity $\tau$ as total tokens per completed task and let $Q\in[0,1]$ be task quality under a fixed judge protocol. We use two efficiency metrics:定义 Token 强度 τ 为每完成任务的总 Token 数，Q 为评测协议下的任务质量。我们使用两个效率指标...

$$\eta_{\mathdollar}\;=\;\frac{Q}{C}\quad\text{(quality per dollar)},\qquad\mathrm{CPM}\;=\;\frac{Q\cdot 10^{6}}{\tau}\quad\text{(task-completions per million tokens)}.$$

(5)

Definition 1 (Token maxing). A development trajectory $\{(Q_{t},\tau_{t})\}_{t}$ exhibits token maxing if token intensity grows, $\tau_{t+1}>\tau_{t}$, while marginal quality per token declines, $\frac{Q_{t+1}-Q_{t}}{\tau_{t+1}-\tau_{t}}<\frac{Q_{t}}{\tau_{t}}$ — i.e., each release buys quality at a worse token exchange rate than the system’s running average.定义 1（Token 最大化）：开发轨迹 {(Qt,τt)}t 若满足 τt+1 > τt 且边际质量每 Token 递减，即 (Qt+1-Qt)/(τt+1-τt) < Qt/τt，则表现为 Token 最大化。即每次发布换来的质量提升，不如系统平均水平。

Token maxing is individually rational for a team judged on benchmark quality and collectively expensive for the organization paying per token. It is also self-reinforcing under falling prices: when $p$ drops, the budget-constrained optimum shifts toward higher $\tau$, and total spend $N\cdot C$ can rise even as $p\cdot\tau$ per task falls—Jevons’ coal, restated for tokens [1, 24]. The escape is not cheaper tokens but a higher $\mathrm{CPM}$: doing the same work with fewer tokens. Equation (2) says four of the five input terms, plus retries, are code, not model. That places the escape route squarely in the harness.Token 最大化对追求基准分的团队是理性的，对付钱的组织则是昂贵的。在降价环境下，它会自我强化：当 p 下降，预算约束下的最优解倾向于更高 τ，导致总支出不降反升。逃离路径不是更便宜的 Token，而是更高的 CPM：用更少 Token 做同样的工作。Eq.(2) 表明，五个输入项中的四个，加上重试，全是代码，不是模型。逃离路径就在 Harness 里。

Let a fleet run $N$ tasks per month across models $m\in\mathcal{M}$ with mix weights $w_{m}$. Monthly spend is $\sum_{m}w_{m}N\,C_{m}$. A model-side optimization improves one $C_{m}$; a routing policy improves the mix $w$; a harness improvement multiplies every $C_{m}$ by a factor $(1-s_{m})$ simultaneously—and keeps multiplying when the model set changes, because it is implemented above the model API. Empirically (Section 6.2) $s_{m}\in[0.33,0.61]$ across all six models with no exceptions, which is what makes the harness the unusual asset in the stack: the one whose returns are indifferent to which vendor wins the model race.设月度任务量为 N，模型组合权重为 wm，月支出为 ∑wm N Cm。模型侧优化只改进一个 Cm；路由策略改进 wm；而 Harness 改进将每个 Cm 乘以系数 (1-sm)。由于它在模型 API 之上实现，当模型集更换时，它依然有效。实证（第 6.2 节）显示 sm 在 0.33 到 0.61 之间，无一例外。这是 Harness 的独特资产属性：无论哪家模型厂商赢，它都赚钱。

In Writer’s architecture the harness is the runtime between the application and any foundation model: it owns (i) context assembly—system prompt, conversation state, retrieval payloads from document and knowledge-graph grounding; (ii) the tool layer—native tools and external connectors via the Model Context Protocol [23], including schema exposure and call mediation; (iii) workflow execution—multi-step Playbooks run end to end; (iv) delegation—spawning scoped sub-agents and merging their results (absent from the baseline loop); and (v) observability—a trace shim that records prompt tokens, completion tokens, tool events, and wall-clock (duration_seconds) for every turn of every run. The trace shim is what makes this paper’s accounting possible: cost is computed at report build by applying a pinned public price table to the recorded token counts, so both arms are priced identically. The same layer is the enterprise control plane: the trace shim that meters tokens is also the audit trail, the progressive tool disclosure that saves tokens is also tool governance, and deterministic workflow execution is what makes agent behavior reviewable—efficiency and control are properties of one component, which is why the harness sits at the core of everything else the platform does.Writer 的架构中，Harness 是应用与基础模型间的运行时：它拥有 (i) 上下文组装；(ii) 工具层（原生工具及 MCP）；(iii) 工作流执行；(iv) 委派（生成子智能体）；(v) 可观测性（记录 Token、工具事件、耗时）。追踪 shim 让本文的核算成为可能：成本在报告生成时通过固定价格表计算。同一层也是企业控制平面：计量 Token 的 shim 也是审计跟踪，渐进式工具披露既省 Token 又做治理，确定性工作流让行为可审查——效率与控制是同一组件的属性。

The harness is model-agnostic by construction: the executing model is a configuration value (model_name), which is precisely what permits the controlled swap in Section 5—six models, two orchestration layers, one task set, one judge panel. It is the runtime behind Writer’s Action Agent, whose public results (61% on the hardest GAIA level; the top overall score on the CUB computer-use benchmark) situate the production system this evaluation instruments [26].Harness 是模型无关的：执行模型只是一个配置值。这正是第 5 节受控实验的前提——6 个模型、2 种编排、1 套任务。它是 Writer Action Agent 后面的运行时，该系统在 GAIA 和 CUB 基准测试中表现顶尖。

The baseline is the conventional agent loop as previously deployed in production, frozen on 2026-06-07. Its design is the industry default and worth stating concretely, because each element is a token-economics decision made by omission: a monolithic system prompt of roughly 49 KB replayed on every turn; tool invocation parsed by regular expressions from XML emitted in the text stream; destructive middle-truncation when context overflows; per-model prompt tuning; and waits implemented as polling. The Writer Agent Harness, evaluated at its intended general-availability configuration, replaces each with an explicit contract: one execution path for any of the six models; native tool calling only (the XML path was deleted, not ported); non-destructive structured compaction in place of truncation; early streaming with live tool-call status; cancellation and retry as first-class states; and sub-agents—delegating a sub-task to one or more spawned agents with scoped context and merging their results—a capability the baseline loop does not possess at all.基准是 2026-06-07 冻结的常规生产循环。其设计是行业默认做法，每个元素都是通过“省略”做出的经济决策：约 49KB 的单体系统提示词每轮重发；XML 解析工具调用；上下文溢出时破坏性截断；按模型调优提示词；轮询等待。Writer Agent Harness 替换了这些：统一执行路径；仅原生工具调用；非破坏性结构化压缩；早期流式传输与实时状态；取消与重试作为一等状态；子智能体委派——这是基准循环完全不具备的能力。

Mapping the observed deltas onto Eqs. (2) and (4), the savings must come from the terms the harness rewrote. The design goal can be stated in one sentence: maximize the fraction of tokens that are (a) cached, (b) decision-relevant, and (c) spent inside committed, recoverable work—and enforce all three with structure rather than model behavior. Six mechanism families implement it.节省必须来自 Harness 重写的项。设计目标一句话：最大化 (a) 缓存，(b) 决策相关，(c) 投入在可恢复工作中的 Token 比例——并用结构而非模型行为来强制执行。六大机制实现之。

Recent work argues the harness should be a first-class object of design and evaluation [27], and cross-harness measurement finds that agent capability is a property of the model–harness configuration rather than of the model alone [28]. Our claim here is narrower and economic: among widely used agent systems, token economics is nowhere a first-class, published contract. Table 1 summarizes the landscape on the mechanism families of Section 4.3, assessed from public documentation and from a source-level study of caching behavior conducted during this harness’s design.最近工作认为 Harness 应成为设计与评估的一等对象。我们的主张更窄且经济：在主流 Agent 系统中，Token 经济学从未成为一等发布的契约。表 1 总结了基于公开文档的机制评估。

The experiment is a paired swap. Both arms execute the same 22 tasks on the same six models with the same judges and the same pinned price table; only the orchestration layer differs. The baseline arm is the conventional production loop invoked via model_name override, run once and frozen on 2026-06-07 as the reference. The harness arm is the Writer Agent Harness at its intended general-availability configuration. Because prompts, models, and judges are held fixed, any systematic difference in tokens, cost, latency, or quality is attributable to the harness.实验是配对交换。两组在 6 个模型上执行相同的 22 个任务，使用相同评测器和价格表；仅编排层不同。基准组是 2026-06-07 冻结的常规生产循环。Harness 组是 Writer Agent Harness。由于提示词、模型、评测器固定，任何系统性差异均归因于 Harness。

The evaluation uses $n=22{}$ locked-baseline prompts, audited for capability coverage and for defensible pass/fail criteria before either arm ran. The set spans nine capability areas that mirror production usage of an enterprise agent platform: Model & System Awareness (MSA: identity, scope, safe refusal), Grounding & Retrieval (GDR: document and knowledge-graph grounding, web search, citation), Content Generation (CNG), Playbooks (PLY: predefined multi-step workflows executed end to end), MCP tool use (MCP: discovering and calling external connectors), Presentations (PRN: slide generation, an early-stage capability), Voice (VOX: brand voice and custom instructions), Image Analysis & Generation (IMG), and—under the harness only—Sub-agents (delegated work, net-new). Tasks include multi-turn cases (e.g., a three-turn Medicare grounding dialogue) and long-horizon cases (multi-step research synthesis).评估使用 22 个锁定基准提示词，审计了能力覆盖和通过/失败标准。领域涵盖：模型与系统意识（MSA）、接地与检索（GDR）、内容生成（CNG）、工作流（PLY）、MCP 工具使用（MCP）、演示（PRN）、语音（VOX）、图像分析（IMG）以及仅在 Harness 下的子智能体（SUB）。任务包含多轮对话和长周期研究合成。

Table 2 lists the six models: two frontier generalists, one fast-tier model, two open-weight candidates, and Writer’s platform-native Palmyra X6. The spread is deliberate—the central claim is about a layer above the model, so it must be tested across weight classes and vendors, not on a single flagship.表 2 列出六种模型：两款前沿通用型，一款快速响应型，两款开源权重型，还有 Writer 自家的 Palmyra X6。选得杂，是有意的。核心论点在于模型之上的那层架构。所以不能只测旗舰款，得看不同量级、不同厂商的表现。

Quality. Headline quality is task-completion scored by an LLM-judge panel against the audited pass/fail criteria [20]. Secondary judges score coherence, communication quality, and hallucination; these are reported per-prompt and did not headline any decision. Tokens. Prompt and completion tokens are recorded per turn by the trace shim; token intensity $\tau$ is the per-task total. Cost. Computed at report build as $(\text{prompt tokens}\times p_{\mathrm{in}})+(\text{completion tokens}\times p_{\mathrm{out}})$ using a pinned OpenRouter price table, identically for both arms. Latency. Wall-clock from first token to terminal event (duration_seconds), median across runs. UX. A hand-scored qualitative rubric (streaming responsiveness, tool-call transparency, citation rendering, error/cancel handling, multi-turn coherence) reported separately from automated metrics.质量。以 LLM 裁判组根据审计标准 [20] 判定的任务完成度为准。辅以连贯性、沟通质量和幻觉评分，这些只在单次提示词中记录，不作为决策依据。Token。由追踪垫片记录每轮的输入输出 Token。Token 强度 τ 为单任务总量。成本。在报告生成时，按 OpenRouter 冻结价格表计算：(输入 Token × 进价) + (输出 Token × 出价)。两套框架算账方式一致。延迟。从第一个 Token 到结束事件的墙上时间（秒），取多次运行的中位数。UX。人工评分，包含流式响应、工具调用透明度、引用呈现、错误处理和多轮连贯性，与自动化指标分开报告。

At $n=22{}$, quality deltas are directional, not statistically significant, and we treat them as such throughout: the headline $+0.03$ is reported as parity, not improvement. Cost, token, and latency deltas are large, uniform in sign across all 22 prompts and all six models, and therefore decisive at this sample size in a way the quality deltas are not. Candidate-model failures on advanced features are expected and scored, not excluded; model-agnosticism is a claim about the execution path, not a promise of uniform capability.样本 n=22。质量差值仅具方向性，统计学上不显著。所以统一视为持平，+0.03 的提升不算改进。成本、Token 和延迟的差值很大，且在 22 个提示词和 6 种模型中表现一致。在这个样本量下，这些指标具有决定性。模型在高级功能上失败是预料之中，照样记分，不予剔除。模型无关性是架构路径的声明，不是对能力的承诺。

Table 3 and Figure 3 give the blended result. Replacing the baseline with the harness removes 38% of token intensity and 41% of cost while quality stays at parity—which is the operational definition of escaping token maxing: the token exchange rate improved instead of degrading. In efficiency terms, a dollar buys 82% more completed quality and a million tokens completes 68% more tasks. The hand-scored UX rubric moved the same direction for the mechanical reasons one would expect from the architecture: earlier streaming, live tool status instead of a spinner, clean cancellation.表 3 和图 3 展示了结果。换上新框架，Token 强度减 38%，成本减 41%，质量维持不变。这就是所谓的“逃离 Token 上限”：Token 兑换率提升了，没降级。换算效率，一美元能多买 82% 的质量，一百万 Token 能多做 68% 的任务。人工 UX 评分也同步向好：流式传输更快，工具状态实时显示，取消操作更干脆。

Table 4 and Figure 4 break the effect out per model. Two observations. First, uniformity: six models, five vendors, three weight classes, and not one exception—every model’s cost falls by at least a third. This is the signature of a layer-level effect: if the savings came from any model-specific behavior, the spread would show it. Second, magnitude relative to model choice: under the baseline, moving from the most expensive model (Palmyra X6 at $0.25) to the cheapest (Qwen 3.6 at $0.16) saves 36%; keeping any model and adopting the harness saves 33% to 61%. On this workload, the orchestration layer is a larger cost lever than the model menu—which inverts where optimization attention usually goes.表 4 和图 4 展示了单模型表现。两点观察。第一，统一性：六种模型、五家厂商、三个量级，无一例外。成本全降至少三分之一。这是层级效应的特征：如果是因为模型特异性，分布不会这么齐。第二，相对于模型选择的量级：基准下，最贵模型（Palmyra X6，$0.25）换成最便宜（Qwen 3.6，$0.16）省 36%；保持模型不动，换上新框架省 33% 到 61%。在这个工作负载下，编排层比选模型更能省钱。这颠覆了常规的优化重心。

The largest relative gains land on the fast tier: Flash 3.5 drops 61% in cost and 55% in latency. This is consistent with the decomposition in Eq. (2): for small, cheap models, harness overhead (replayed history, broadcast schemas) is a larger share of the total bill, so removing it removes proportionally more.快速 tier 收益最大：Flash 3.5 成本降 61%，延迟降 55%。这符合公式 (2) 的拆解：对于小而便宜的模型，框架开销（历史重放、广播模式）占大头，去掉它，省得就多。

Figure 5 plots all 48 capability$\times$model cells (eight legacy capabilities $\times$ six models; the full matrix is Appendix A). The mass sits on or above the diagonal: 30 cells improve, 11 are flat, 7 regress. The regressions are not random. Every one occurs on Flash 3.5, Qwen 3.6, or GLM 5.1, and six of the seven fall in capabilities that exercise orchestration hardest—MCP tool use (Qwen $-0.15$, GLM $-0.06$, Flash $-0.04$), Playbooks, and Presentations. The frontier models and Palmyra improve most in exactly those categories (MCP: Sonnet $+0.10$, Palmyra $+0.10$; GDR: Sonnet $+0.10$, Palmyra $+0.12$). The same richer harness that a strong model converts into quality, a weaker model experiences as load.图 5 绘制了 48 个单元。重心在对角线上方：30 升，11 平，7 降。退步不随机。全发生在 Flash 3.5、Qwen 3.6 或 GLM 5.1 上。七个里有六个是编排重头戏：MCP 工具使用、Playbooks 和演示文稿。前沿模型在这些领域提升最明显。框架越强，强模型转化出质量，弱模型却感到吃力。

Collapsing each model to its mean across the eight scored capabilities gives a single number for how much quality it extracted from the swap: Palmyra X6 $+0.079$, Sonnet 4.6 $+0.073$, Gemini 3.1 $+0.050$, GLM 5.1 $+0.028$, Flash 3.5 $+0.010$, Qwen 3.6 $-0.031$. Plotted against the baseline (Figure 6), the relationship is nearly linear ($r=0.99$; six points, so suggestive rather than conclusive). We call the slope harness leverage: the rate at which a model converts orchestration structure into quality. The economic consequence is asymmetric and useful: efficiency gains are unconditional, quality gains are earned by capability. A weak model under the harness still gets its 44–61% cost cut; it simply does not also get better. This decouples the two reasons to upgrade a harness and lets each be priced on its own.将每个模型在八项能力上的平均分折算成一个数：Palmyra X6 +0.079，Sonnet 4.6 +0.073，Gemini 3.1 +0.050，GLM 5.1 +0.028，Flash 3.5 +0.010，Qwen 3.6 -0.031。与基准对比（图 6），关系近乎线性。我们称之为框架杠杆：模型将编排结构转化为质量的比率。经济后果不对称：效率提升是无条件的，质量收益靠能力换取。弱模型用新框架，成本照样降 44–61%，只是质量没变好。这让效率和质量的升级理由解耦，可以分别定价。

The harness’s one genuinely new capability—delegating work to spawned sub-agents—lands above a usable reliability threshold only on the two strongest models (Table 5: Palmyra X6 at 0.86, Sonnet 4.6 at 0.85), degrades on Gemini 3.1 (0.70) and GLM 5.1 (0.58), and is not yet dependable on the fast tier (0.42–0.45). This is the harness-leverage result in its sharpest form: an orchestration feature carries a capability floor, below which exposing it produces failures rather than function. Delegation also has a distinctive token-economics profile—sub-agent context is scoped, so delegated tokens are spent in a bounded side-loop rather than inflating the parent context—the context-firewall contract of Section 4.3: a capped summary returns, citations ride a sidecar, and the parent loop never pays for the child’s exploration.框架真正的新能力——委派工作给子代理——只有在两个最强模型上才跨过可用门槛。Gemini 3.1 和 GLM 5.1 退步，快速 tier 尚不可靠。这是框架杠杆最直观的表现：编排功能有能力门槛，低于门槛只会产生故障，而非功能。委派的 Token 经济学也不同：子代理上下文受限，委派的 Token 在侧回路消耗，不膨胀父上下文。这实现了 4.3 节的上下文防火墙：返回汇总，引用走侧车，父循环不为子循环的探索买单。

Four exemplar prompts show the distribution behind the aggregates. The three-turn Medicare grounding dialogue improves from 0.60 to 0.90—the single largest quality jump in the set—under a harness whose retrieval shaping sends less, better-selected evidence, consistent with the finding that long, noisy contexts degrade attention [17]. Identity/refusal holds at 0.90/0.90 while its cost halves ($0.04 $\rightarrow$ $0.02): safety behavior at half price. Contract Q&A moves $+0.07$ (0.75 $\rightarrow$ 0.82). The most expensive task in the set, multi-step research synthesis, drops from $0.61 to $0.33 ($-46\%$) but regresses in quality (0.80 $\rightarrow$ 0.60)—the one place the aggregate parity conceals a real trade, driven by the smaller models, and the reason the release recommendation (Section 7) holds candidates back pending a fix rather than shipping the regression. Secondary judges move with the headline: coherence 0.85 $\rightarrow$ 0.88, communication 0.79 $\rightarrow$ 0.80, hallucination clean in both arms.四个示例展示了总量背后的分布。三轮 Medicare 溯源对话质量从 0.60 升至 0.90，是所有任务中提升最大的。框架优化了检索，送入的证据更少、更准。这印证了长而嘈杂的上下文会分散注意力 [17]。身份/拒绝任务在 0.90/0.90 持平，成本减半：安全行为，半价实现。合同问答升 0.07。最贵的任务——多步研究综合——成本从 $0.61 降至 $0.33，但质量退步（0.80 → 0.60）。这是唯一一处总量持平掩盖的真实折损，由小模型驱动。这就是为什么发布建议（第 7 节）要扣住候选模型，先修补再发布。辅裁判指标同步：连贯性 0.85 → 0.88，沟通 0.79 → 0.80，幻觉两边都干净。

The results separate cleanly along the framing of Section 3. The terms of Eq. (2) that are pure code—history, schemas, retrieval size, retries—moved for every model, because code executes the same way regardless of which model reads its output. The quality terms are mediated by the model’s ability to exploit structure, and they moved with capability. One layer, two currencies: the harness sets the price of work unconditionally, and sets the ceiling of work jointly with the model. The layer that sets the price is also the layer an enterprise governs by; efficiency and control concentrate in the same component.结果按第 3 节的框架清晰分割。公式 (2) 中纯代码的部分——历史、模式、检索大小、重试——对所有模型都变了，因为代码执行方式不随模型而变。质量项由模型利用结构的能力调节，随能力变动。一层架构，两种货币：框架无条件设定工作价格，与模型共同设定工作上限。设定价格的层级，也是企业治理的层级；效率和控制权集中在同一个组件。

This reframes a common procurement instinct. Teams comparing $/Mtok across vendors are comparing $p$; the bill is $p\times\tau$, and $\tau$ belongs to the harness. On this workload the harness moved the bill more than the entire spread of the model menu did (Section 6.2). An organization that rents its orchestration layer has outsourced the variable it controls most.这重构了采购逻辑。团队比较厂商间的 $/Mtok，其实是在比 pp；账单是 p×τ，而 τ 在框架手里。这次测试，框架对账单的影响力超过了所有模型差价的总和。租用编排层的组织，等于外包了自己最能掌控的变量。

Per-task deltas understate the object of interest, which is fleet spend. At the blended rates measured here, an organization running one million agent tasks a month pays $210k under the baseline loop and $120k under the harness—$1.08M a year from an orchestration change alone (Figure 7). Three properties make this saving unusual among AI optimizations. It is model-portable: implemented above the API, it survived all six models here and, by construction, applies to models that do not exist yet. It is volume-linear: it grows with exactly the quantity—agentic task volume—that is growing fastest in enterprise deployments. And it stacks: per-token price declines [24], routing [10, 11], and prompt-level compression [9, 13] all multiply against it rather than substituting for it. Latency compounds the same way: 1.8$\times$ faster per task is also 1.8$\times$ the throughput per unit of orchestration infrastructure, and shorter loops are cheaper to retry.单任务差值低估了车队支出。按测得的综合费率，每月一百万任务，基准循环需 $210k，框架只需 $120k——仅凭编排变更一年省 $1.08M（图 7）。这种节省在 AI 优化中很少见。它模型可移植：在 API 之上实现，六种模型通用，未来模型也适用。它规模线性：随企业部署中增长最快的代理任务量同步增长。它可叠加：Token 单价下降、路由和提示词压缩，全都能乘上去，而不是替代它。延迟也一样：单任务快 1.8 倍，意味着编排基础设施的吞吐量也高 1.8 倍，且短循环重试成本更低。

Definition 1 suggests the managerial fix is a measurement fix. Teams that report quality alone will token-max, because tokens are someone else’s line item; teams that report $\mathrm{CPM}$ or $\eta_{\mathdollar}$ cannot. The swap studied here moved $\mathrm{CPM}$ from 54.9 to 92.0 completions per million tokens—the direction opposite to the industry trajectory—while quality held. We suggest CPM belongs next to quality in every agent release gate, for the same reason performance-per-watt sits next to performance in chip design: it is the number that predicts the bill.定义 1 暗示管理上的修复是测量上的修复。只报质量的团队会触碰 Token 上限，因为 Token 是别人的开支；报 CPM 或 η$ 的团队则不会。本次切换将 CPM 从 54.9 移至 92.0，与行业轨迹背道而驰，同时质量未降。建议在每个代理发布门槛中加入 CPM，就像芯片设计中性能/瓦特必须紧跟性能一样：它是预测账单的数字。

The regression pattern (all seven on smaller models, concentrated in MCP/Playbooks) and the sub-agent floor (Table 5) argue that orchestration features are not free to expose. A harness is a contract the model must be strong enough to honor. Two practical consequences follow. First, harness capabilities should degrade gracefully by model tier—scoping down tool catalogs and disabling delegation below the floor—rather than presenting one interface to all models. Second, routing research [10, 11] should route on feature demand, not just prompt difficulty: a request that will exercise sub-agents belongs on Palmyra X6 or Sonnet 4.6 regardless of how simple its text looks, while a grounded Q&A request can take the 61%-cheaper fast tier with no quality penalty (Figure 5, GDR row: every model improves).退步模式（全在小模型，集中在 MCP/Playbooks）和子代理门槛（表 5）证明编排功能不是免费的。框架是模型必须有足够能力去遵守的契约。两点后果。第一，框架功能应随模型等级优雅降级——低于门槛就缩减工具目录、禁用委派，而不是对所有模型展示同一界面。第二，路由研究应按功能需求路由，不只是按提示词难度：涉及子代理的请求，无论文本多简单，都应去 Palmyra X6 或 Sonnet 4.6；而溯源问答请求可以走便宜 61% 的快速 tier，质量无损（图 5，GDR 行：所有模型都提升）。

Read as a deployment decision, the evidence supports: general availability on the two models above the sub-agent floor (Palmyra X6, Sonnet 4.6); holding the open-weight candidates pending the multi-step-research fix; and leading external claims with efficiency—which is uniform and decisive—rather than quality, which is directional at $n=22{}$. We record this here because it is the honest reading of Table 3: the temptation in industry reporting is to headline $+0.03$ quality; the defensible headline is $-38\%{}$ tokens at parity.作为部署决策，证据支持：在跨过子代理门槛的两个模型（Palmyra X6, Sonnet 4.6）上全面发布；开源候选模型暂缓，等待多步研究修复；对外宣传以效率为主——这是统一且决定性的——而不是质量，质量在 n=22 下仅具方向性。这是表 3 的诚实解读：行业报告倾向于标榜 +0.03 的质量；可辩护的标题是：同等质量下，Token 减少 38%。