Control the Harness, Control the Cost: Routing and Governing AI Coding Agents in the Enterprise
Arian Abbasi, Alan Aqrawi, Ted Kwartler
cs.AI · cs.CR
Quick verdict
Claims cache-safe routing saves 14–21% of enterprise agent model spend; cache arithmetic is sound, but headline rests on 46 author-written synthetic scenarios.
Summary
The paper argues that in enterprise AI coding agents, the harness — not the price sheet alone — determines model spend, because it selects models, manages the prompt cache and launches subagents. The authors derive a cache-aware cost model: an equation for when a mid-task model switch repays its cache writes (8–291 requests at Anthropic's September 2026 list prices), and a crossover showing the top-priced model (Fable 5.1) becomes cheaper than Opus 5 on long tool-heavy sessions thanks to half-price cache reads; repricing ~10,000 real sessions from two public corpora shows 51–80% of $100+ sessions are cheaper on the top model. They specify a router that moves work only at cache-safe points (session start, side lanes, subagent launch), gated by Jev, a fast calibrated classifier over a bring-your-own taxonomy, and emulate a 10,000-seat enterprise from 46 author-written scenarios expanded with measured behaviour, reporting 14–21% spend recovery ($3.3–5.0M/year). Key limitations: scenarios and ground-truth labels come solely from the authors, classifier accuracies are 63–82%, 12.5% of moved turns are false downgrades, code and catalogue are unreleased, prices come from one vendor on one date, and the engineer-day spend assumption triples the vendor-reported average.
Review complete
Assessment at a glance
6
risk flags identified
● high 2
● medium 3
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Because one user turn is k+1 requests over a model-scoped prompt cache, per-request chat routing fails inside harnesses, and a mid-task switch pays back only after 8–291 requests depending on the model pair (Eq. 1).
- On long tool-heavy sessions the highest-priced model (Fable 5.1) costs less than the next tier (Opus 5) due to half-price cache reads — true for 51–80% of real sessions over $100 in two public corpora.
- A router that moves work only at cache-safe points (session start, side lanes, subagent launch), gated by a fast calibrated classifier, recovers 14–21% of model spend for a 10,000-seat enterprise ($3.3–5.0M/year at Anthropic's 21 Sept 2026 list prices).
- User behaviour, especially pauses past the 5-minute cache lifetime (17–27% of turns), moves the saving as much as classifier quality does (7.4% to 24% across pause regimes).
- Strong third-party models cost 0.12–0.80x Opus 5 on identical sessions, but harnesses bound to one vendor cannot reach them, making the control plane, not the price sheet, the decision that governs vendor reach.
Strengths
Cache cost arithmetic is derived, not asserted, and is independently checkable
Eq. 1 gives the payback threshold for mid-task switches (8–291 requests by model pair) directly from Table 1 cache multipliers; Table 2 and Fig. 2 make the crossover falsifiable arithmetic: Fable 5.1 reads cache at $0.25/M vs Opus 5's $0.50/M, breaking even above 65k tokens at 18 tool steps.
Counterintuitive, actionable headline finding grounded in real sessions
The result that the top-priced model is cheaper for long tool-heavy work is validated by repricing ~10,000 real Claude Code sessions from two public corpora (TraceLab, SWE-chat): 51–80% of $100+ sessions cost less on Fable 5.1, and spend concentrates there (top 1% of sessions carry 31–53% of spend, Fig. 3).
Unusually candid limitations section
Extensive, well-attributed sensitivity analysis
Governance-aware system design
Weaknesses
Evaluation corpus is self-authored and self-labelled
Classifier quality is weak exactly where routing decisions live
Nothing is released; auditability is asserted, not demonstrated
Headline is hostage to one vendor's price structure on one date
Core dependency is a closed, vendor-benchmarked hosted classifier
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper addresses enterprises that buy rather than build agentic coding harnesses (Claude Code, Codex, Gemini CLI and peers) and thereby inherit vendor defaults: model choice, prompt-cache handling, subagent placement. It claims five contributions (Sec. 1): (1) a cost model of agentic turns under model-scoped caching, yielding a payback rule for mid-task switches (Eq. 1: 8–291 requests on Anthropic's 21 September 2026 sheet) and a crossover result — on long tool-heavy sessions the top-priced model (Fable 5.1, $0.25/M cache reads) undercuts Opus 5 ($0.50/M) above 65k tokens at 18 tool steps (Table 2), confirmed by repricing ~10,000 real sessions (51–80% of $100+ sessions cheaper on the top model, Fig. 3); (2) a cache-safe router acting only at session start, side lanes and subagent launch (rules R1–R10); (3) a 10,000-seat emulation reporting 14–21% spend recovery ($3.3–5.0M/year); (4) a 20-harness coupling matrix (Table 8); (5) lock-in repricing (0.12–0.80x Opus 5, Fig. 10), a control-plane architecture and a sovereignty ladder (Table 7).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
How much can routing save on enterprise AI coding agent costs?
The paper estimates 14–21% of model spend — $3.3M to $5.0M per year for a 10,000-seat enterprise at Anthropic's September 2026 list prices — by moving work only at cache-safe points. Treat this as an upper-leaning illustration: it comes from simulated scenarios written by the authors, and falls to roughly 5% under a uniform cache-pricing regime or about 7% for pause-heavy user populations.
Why is the most expensive AI model sometimes cheaper overall?
Vendors bill re-read conversation context (cache reads) separately per model. Anthropic's top model reads cached context at $0.25 per million tokens versus $0.50 for the next tier. In long, tool-heavy sessions where every tool step re-sends the conversation, those discounted re-reads dominate, and above roughly 65k tokens of context (at 18 tool steps per turn) the pricier model wins. The authors found 51–80% of real sessions costing over $100 fall in that regime.
Why can't you just switch to a cheaper model mid-task to save money?
The prompt cache belongs to one model. Switching discards it and forces the full conversation to be rewritten at the cache-write premium, and switching back repeats the cost. The paper derives payback thresholds of 8 to 291 requests depending on the model pair, which is why its router only switches at session start, in side conversations, or when launching subagents.
Is the paper's simulation reproducible?
Only partially. The routing taxonomy and cost-model constants are published in the appendices and the two session corpora it draws on are public, but the authors do not release their simulator, the 46-scenario catalogue, or the cached classifier labels, so no external party can regenerate the headline numbers; the classifier itself is a closed hosted model.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures were available only as embedded text with plotted numbers; visual inspection of curves (Figs. 1–3, 9–13) was not possible.
- The paper's September 2026 price sheets, model names (Fable 5.1, Opus 5, Jev, GPT-6 Astra) and many 2026-dated citations could not be independently verified within this review.
- Table 8 (20-harness coupling matrix) was extracted as dense text; some column-to-value alignment was inferred from context.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.