cs.AISep 24, 2026

Control the Harness, Control the Cost: Routing and Governing AI Coding Agents in the Enterprise

Authors: Arian Abbasi, Alan Aqrawi, Ted Kwartler

Organizations: Accenture Responsible AI · Accenture Americas Advanced AI Practice · Harvard Extension School, Harvard University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Harnesses, the products that run AI coding agents, are multiplying, and enterprises are rolling them out to their employees: what started as pilots with a few hundred seats is scaling to tens of thousands. Most enterprises do not build these harnesses but buy them from large vendors, such as Anthropic's Claude Code or OpenAI's Codex. A harness decides which model answers, what the model reads, how the prompt cache is used and which subagents run, so it picks the rate on the price sheet and sets the volume bought at it. Enterprises that keep a proprietary or untuned harness at its defaults inherit these choices and their bill. We build a fast, customisable router in which Jev, a classifier with calibrated probabilities, labels every prompt against a bring-your-own taxonomy of agentic requests. Because one user turn is many requests over a prompt cache that belongs to one model, the router moves work only where no running conversation has to rebuild its cache: at session start, in side lanes and at subagent launch. From the price sheet we derive when a mid-task switch pays back, and a crossover: on long tool-heavy sessions the highest-priced model costs less than the next tier, as repricing about 10,000 real sessions from public datasets confirms. In an emulated enterprise of 10,000 seats with user behaviour taken from these datasets, the router recovers 14 to 21% of model spend at Anthropic's list prices of 21 September 2026, $3.3M to $5.0M a year. The paper also maps the risks across twenty harnesses, prices the dependence on one vendor's models, and proposes a control plane that enterprises can run from within, starting now, with a ladder for deciding later whether to own the harness.

Figures & tables

VendorCache readCache writeLifetimeMinimum prefix
Anthropic [ 11 , 10 ]0.025 ×\times (Fable 5.1), 0.1 ×\times others1.25 ×\times (5 min), 2 ×\times (1 h)5 min or 1 h512 to 4,096 tokens
OpenAI, GPT-5.6 and later [ 12 , 13 ]0.1 ×\times1.25 ×\times30 min1,024 tokens
Google, Gemini [ 14 ]0.1 ×\timesinput price plus a storage fee per hour (explicit cache)not stated (implicit caching on by default)2,048 to 4,096 tokens
Table 1: Prompt-cache price rules by vendor (Anthropic’s of 21 September 2026, OpenAI’s and Google’s of 23 September 2026). Multipliers are of the model’s input price.
Tool steps per user turnContext above which Fable 5.1 is cheaper than Opus 5
3385k
6211k
12105k
1865k
3032k
Table 2: Context above which Fable 5.1 is cheaper than Opus 5 for one warm turn, by tool steps per turn (assumptions as in Figure 2 ).
QuantityValue
Scenarios / turns / job types46 / 109 / 11
Classifier accuracy: action / domain / complexity / history74% / 80% / 63% / 82%
False downgrades among moved turns / held in place (of routable)12.5% / 42%
As-is spend, 10,000 seats1,974kamonth,1,974k a month, 23.7M a year
Routed, today’s labels1,701kamonth(1,701k a month ( -$ 13.8%)
Routed, perfect labels1,558kamonth(1,558k a month ( -$ 21.1%)
Table 3: Case-study summary (list prices of 21 September 2026; 10,000 seats; live classifier labels, cached).
Case (population, same labels)As-is / monthRouted / monthSaving
Base case$1.97M$1.70M13.8%
Ledger rebase at task boundaries$1.97M$1.70M13.9%
One-hour cache lifetime (as-is and routed)$1.93M$1.60M17.1%
No Fable 5.1 licence (as-is on Opus 5, router capped at Opus 5)$2.01M$1.97M2.3%
Everyone as-is on Opus 5 (Fable 5.1 licensed for the router)$2.01M$1.70M15.6%
Confidence threshold 0.7 (base 0.8)$1.97M$1.78M9.8%
Table 4: Sensitivities (population, monthly, same labels).
Always-loaded prefix per requestFable 5.1Opus 5Sonnet 5
15k: lean default (system prompt and core tools)$0.90$1.51$0.60
25k: default plus a 10k instruction and skills bundle$1.49$2.52$1.01
55k: observed power-user configuration with plugins$3.29$5.54$2.22
159k: the same if deferred tool schemas were loaded eagerly$9.50$16.02$6.41
Table 5: Cost per session of the always-loaded prefix (ten turns of eighteen tool steps, warm cache; list prices of 21 September 2026). Every token in the prefix is re-read on every request; the 55k row is a measured power-user configuration.
10,000 returns at 300k contextOn the returnEach later requestReturn vs re-send
Re-send the full history (default)$18,750$1,500+0%
Prune to 30k first (clear old reasoning and tool results)$1,875$150-90%
Compact to 30k, summary by Opus 5$20,625$150+10%
Compact to 30k, summary by Sonnet 5$9,375$150-50%
Table 6: Returning to a conversation whose cache has expired: 10,000 returns at 300k tokens of context on Opus 5 (list prices of 21 September 2026; 5-minute cache). Pruning clears old reasoning and tool results without a model call; a summary is written by the named model, which reads the full history uncached. The 30k is 15k of system prompt and tools plus a 15k-token summary. “Each later request” is the cache read of the history on every request after the return.
RungWhat it buysWhat it costs
0. Vendor defaultsNothing to runEvery risk in Section 3 ; the highest bill
1. ConfiguredThe bundle above, model defaults, subagent default, permissionsConfiguration per harness; no measurement
2. ControlledGateway, session-start routing, side lanes via hooks, telemetry, policy floors, taxonomy; the 14 to 21% of Section 5.1A small platform team; classifier hosting; works-council and legal review of the policy
3. OwnedAn internal harness on a vendor SDK or an open-source base with pluggable models, native lanes, one bundle for everyone; other vendors’ models (Figure 10 )A product team; keeping pace with weekly vendor releases; the dependency moves to maintainers
Table 7: The sovereignty ladder.

Explore similar work

CardsList
Sep 14, 2026cs.AI

Harness or Model? Isolating the Harness Effect in Agentic Coding with a Contamination-Controlled Private Suite

An agentic coding system couples a language model to a harness: the tools, prompts and control flow that turn a chat model into an autonomous software engineer. Vendors ship harnesses tuned to their own models, and practitioners assume the vendor-native pairing solves more tasks. We measure that assumption with paired same-model contrasts on a private, contamination-controlled suite of 256 repository and post-cutoff contest tasks. The same 80 tasks ran under claude-agent-sdk and under deepagents on claude-opus-4-8, and under the openai-codex SDK and deepagents on gpt-5.5, with gemini-3.5-flash and deepseek-v3.2 as side cells. 792 of 800 planned runs were graded by an isolated oracle. Neither contrast resolves an average advantage for either harness: -1.25 pp for Opus 4.8 (48.8% vs 50.0%, task-bootstrap 95% CI [-10.0, +7.5]) and +1.25 pp for GPT-5.5 (55.6% vs 54.4%, CI [-4.4, +6.9]). The Opus average combines opposite strata: the native harness trails by 9.0 pp on the 61 repository tasks and leads by 23.7 pp on the 19 contest tasks (label-permutation p = 0.003). The partition was chosen after seeing the data and needs a designed replication. Correctness and completion also separate: 22 of 81 runs cancelled at the wall-clock ceiling had produced a passing patch. Re-priced from raw per-turn usage at frozen list prices, the neutral harness cost 1.3 to 1.6 times as much per solved task on Opus 4.8 and 1.2 times on GPT-5.5. These are observed-usage estimates. On the Anthropic account 58 runs left no usage record, and allocating that spend to either cell would move the Opus ratio between 0.7 and 2.3, so the billed ordering is unresolved. This revision corrects an August 2026 manuscript whose cost figures rested on a usage-semantics defect in our own telemetry (Section 5.1). We release the orchestrator, grading oracle, reanalysis code and derived aggregates. The tasks stay private.

Mohsen Arjmandi

evolutionID GmbH, Germany

Agent HarnessCoding AgentsMean
AskPDF
May 8, 2026cs.AI

Switchcraft: AI Model Router for Agentic Tool Calling

Agentic AI systems that invoke external tools are powerful but costly, leading developers to default to large models and overspend inference budgets. Model routing can mitigate this, but existing routers are designed for chat completion rather than tool use. We present Switchcraft, the first (to the best of our knowledge) model router optimized for agentic tool calling. Switchcraft operates inline, selecting the lowest-cost model subject to correctness. We construct an evaluation framework on five function-calling benchmarks and train a DistilBERT-based classifier, deployed under a latency budget. Switchcraft achieves 82.9% accuracy -- matching or exceeding the best individual model -- while reducing inference cost by 84%, saving over $3,600 per million queries. We find that larger models do not consistently outperform smaller ones on tool-use tasks, and that nominally cheaper models can incur higher total cost due to token-intensive reasoning. Our work enables cost-aware agentic AI deployment without sacrificing correctness.

Sharad Agarwal, Pooria Namyar, Alec Wolman +3

Microsoft Research · Microsoft · Stanford

Agentic InferenceLarge Language Model RoutingLarge Language Model Tool UseInference CostCraft
AskPDF
Jul 8, 2026cs.AI

The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI

Agentic AI development today runs on token maxing: buying capability with tokens -- longer reasoning traces, more turns, wider tool payloads, bigger replayed contexts -- so tokens per task grow faster than task value. Falling per-token prices mask the pattern; total spend rises anyway. We argue the decisive lever against token maxing is the harness: the orchestration layer that assembles context, exposes tools, sequences turns, delegates work, and carries enterprise observability and governance. We isolate it with a controlled swap: 22 locked evaluation tasks, six foundation models (Claude Sonnet 4.6, Gemini 3.1, Gemini Flash 3.5, Qwen 3.6, GLM 5.1, Palmyra X6), changing only the orchestration layer -- a frozen conventional production loop versus the Writer Agent Harness. Holding models constant, the harness cuts blended cost per task 41% (0.21−>0.21->0.12), median wall-clock 44% (48s->27s), and tokens per task 38% (14.2k->8.8k), with task-completion quality at parity (0.78->0.81, directional at this sample size). Efficiency is model-invariant -- every model gets cheaper (33-61%) -- while quality gains are capability-dependent: a model's gain correlates almost perfectly with its baseline strength (r=0.99, n=6), a phenomenon we term harness leverage. Quality per dollar rises 82%; task-completions per million tokens rise from 54.9 to 92.0. On this workload the orchestration layer moved cost per task more than the full spread of the model menu did. We formalize token economics at the orchestration layer (including effective input price under prompt caching), detail the six mechanism families behind the effect -- cache-shape discipline to failure-spend governance -- compare six widely used agent systems on the same axes, and argue the harness is the one component whose efficiency multiplies across every model an organization runs -- present and future.

Muayad Sayed Ali, Aliaksandra Novik, Anji Boddupally +28

Writer, Inc.

Agent HarnessOrchestrationAgentic Inference
AskPDF