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

1 Introduction

• Enterprises are moving coding-agent harnesses from pilots to production, while the share of Codex users sending prompts estimated to require an hour or more of skilled work doubled from 35 to 70% in five months.The introduction cites large enterprise rollouts and more than a million pull requests opened by GitHub Copilot coding agent between May and September 2025.Source: §1 · p. 2
• Purchased harness defaults determine model selection, prompt and cache management, subagents, and logging, shaping both costs and enterprise control.The paper identifies exposure to vendor pricing, closed code, fragmented skills and connectors, and dependence on individual user skill.Source: §1 · p. 2
• Jev classifies each prompt against a bring-your-own taxonomy, and routing moves work only where a running conversation need not rebuild its cache.The router acts at session start, in separate lanes for small tasks, and when launching subagents; Jev operates in under half a second.Source: §1 · p. 2
• 14 to 21% of model spend, or $3.3M to $5.0M yearly, is recovered by routing in an emulated 10,000-seat enterprise.The emulation uses user behaviour measured in real sessions; the paper also maps harness risks and proposes an enterprise-run control plane.Source: §1 · p. 2
• The paper contributes a risk map of twenty harnesses, a model-scoped-cache cost model, a cache-safe router, and an emulation method for estimating costs before telemetry exists.It also reports a case study with sensitivities and proposes a control-plane architecture and sovereignty ladder.Source: §1 · pp. 2–3

2.1 Every tool step re-sends the conversation

• A tool-using turn with k steps makes k+1 API requests, each carrying the conversation so far, so later requests resend more context until compaction or a new conversation.The conversation cache is model-specific; moving the same conversation to another model recomputes it, and cache reads, writes, and expiry affect cost.Source: §2.1 · p. 3
• The price rules vary by vendor: Anthropic’s cache rules are dated 21 September 2026, while OpenAI’s and Google’s are dated 23 September 2026.Table 1 expresses cache-price multipliers relative to each model’s input price.Source: §2.1 · p. 3, Table 1 · p. 4

2.2 Model coupling is a design choice

• Model openness and model coupling are independent: Gemini CLI is open-source but Google-only, while Amp and Copilot CLI are proprietary but support multiple models.Across twenty surveyed harnesses, seven document subagents inheriting the parent model; Chinese-vendor harnesses are mostly open-source and accept compatible endpoints, with CodeBuddy and iFlow CLI as closed exceptions.Source: §2.2 · p. 4

3 Problem statement

• Buying a harness transfers four risks to the enterprise, each set by configuration the enterprise can read today.The supplied passages describe coupling to vendor pricing, code opacity, fragmented portability, and dependence on user skill.Source: §3 · p. 4, §3.1 · p. 4, §3.2 · p. 4

3.1 Coupling ties the enterprise to one vendor’s prices

• A harness bound to one vendor’s models also binds the enterprise to that vendor’s prices, plans, and limits, with coupling ranging from exclusive to multi-model endpoints.The paper gives examples of vendors changing usage-based pricing or subscription limits, making the exposure relevant to procurement and platform teams.Source: §3.1 · p. 4

3.2 Enterprises pay for behaviour they cannot inspect

• Closed harness code can prevent enterprises from inspecting sent content, auditing prompt changes, or verifying cache, retry, and subagent behaviour.The paper identifies several closed harnesses and notes that opacity compounds pricing exposure because the resulting bill cannot be audited.Source: §3.2 · pp. 4–5

3.3 Connectors travel, the control plane does not

• MCP, AGENTS.md, and Agent Skills carry some connectors, instructions, and procedures across harnesses, but model policy, telemetry, hooks, permissions, subagents, and billing remain non-portable.A workforce split across harnesses otherwise faces separate skill bundles, connectors, and governance configurations without shared telemetry.Source: §3.3 · p. 5

3.4 Novices get half the results for the same spend

• Novice-rated sessions reach verified success 15% of the time, versus 28 to 33% for intermediate users and above, and 19% end abandoned versus 5 to 7% for others.The paper reports that harness configuration can reach every user at no marginal cost and maps expertise signals to scaffolds, verification hooks, review gates, and model policy.Source: §3.4 · p. 5

3.5 One layer decides all four risks

• The routing design excludes a productivity axis and reports aggregates only, because per-user cost and productivity metrics raise employee-monitoring concerns under EU and German rules.Data-retention and refusal policies also differ by model, and the routing policy carries those floors.Source: §3.5 · p. 5

4.1 Price the turn, not the request

• A mid-task switch must recoup cache-write costs: Fable 5.1 to Sonnet 5 needs 291 requests with a return switch, versus 46 without.The warm turn reads roughly (k+1)C + Dk/2 cached tokens, writes D, and produces output; moving models turns reads into cold writes. The emulation found naive per-request switching costs more than pinning once misroutes are charged, while the router beats both.Source: §4.1 · p. 6, §4.2 · p. 6

4.3 The top model undercuts the next tier on long sessions

• Above 65k tokens with 18 tool steps, and above 211k with six, Fable 5.1 is cheaper than Opus 5 for a warm turn despite its higher list price.Repricing real sessions found Fable cheaper in 51% of TraceLab and 80% of SWE-chat sessions costing at least $100 on Opus 5; the top 1% of sessions account for 31 to 53% of spend.Source: §4.3 · p. 6

4.4 Route without touching running work

• The cache-safe policy chooses at session start, runs bounded side requests in a separate ledger-seeded lane, and routes subagents when launched.In an 805k-token session, a one-line change costs $0.88 in place on Fable 5.1 versus $0.36 in a Sonnet 5 ledger side lane, including ledger and home refresh.Source: §4.4 · pp. 7–8

4.5 Label with a bring-your-own taxonomy, move only when sure

• The bring-your-own taxonomy labels action, domain, complexity, and context dependence, with governance floors and stakes ceilings constraining cheaper-model routing.The policy holds a turn unless probabilities support both the cheaper model’s capability and the absence of history dependence; corrections inherit the task tier and history dependence.Source: §4.5 · pp. 8–9

4.6 Jev classifies every prompt in under half a second

• Jev scores typed options in parallel and returns calibrated probabilities, letting the policy act only when confidence supports the cheaper tier and history-independence conditions.The interface is classifier-agnostic; a self-hosted encoder is the stated fallback for restricted contexts because hosted classification sends prompts outside the company and Jev is not offered for self-hosting.Source: §4.6 · p. 9

4.7 Reasoning effort is a second dial

• Reasoning effort is a second routing lever: lower effort can save on bounded work, while higher effort can lower cost per completed task on long-horizon work.The passage reports medium effort matched default accuracy at about 70 to 87% of its cost in published knowledge-work runs; on most models, changing effort mid-conversation invalidates the cache.Source: §4.7 · p. 10

4.8 Eight components, from gateway to governance floors

• The proposed control plane combines gateway enforcement, classifier and taxonomy adapters, session state, cache-safe execution modes, counterfactual accounting, user override, and governance floors.The prototype implements components 2 to 6 and 8 as a decision engine and simulator; components 1 and 7 connect it to live harnesses and users.Source: §4.8 · pp. 10–11

4.9 The emulation runs authored tasks with measured behaviour

• The emulation combines 46 authored conversations with 109 turns and 200 variants per conversation, using behavior measured in about 10,000 real Claude Code sessions.It applies live classifier labels and a cache-exact simulator, charges misroutes for observed correction chains, and scales results across 14 population segments under stated model-mix and usage assumptions.Source: §4.9 · pp. 11–12

5.1 A 10,000-seat enterprise recovers 14 to 21%

• 14–21% of model spend is saved, or $3.3M–$5.0M a year, from $23.7M annual as-is spend in the 10,000-seat emulation.Engineers save 13–14%; smaller non-engineering bills save 6–35%.Source: §5.1 · p. 12

5.4 Pauses, the price sheet and tool use move the saving most

• 1.6% savings at half today’s tool steps and tokens per turn rise to 24% at double; lowering the threshold to 0.5 moves 46% more turns but reduces savings to 10%.At threshold 0.9, savings are 14.2%; the authors retain 0.8 as the base rather than tune on this corpus.Source: §5.1 · p. 12, §5.4 · p. 15

5.2 Three quarters of the saving comes from the crossover

• $211k a month comes from choosing the cheapest capable model per session, the largest saving component; side lanes and subagent routing add $64k and $32k.Misroutes and router overhead subtract $51k and $18k monthly, respectively.Source: §5.2 · p. 12
• Fable 5.1 takes 86% of routed model spend while starting one fifth of sessions; $229k monthly savings on moved sessions exceed $23k in costlier moves.Fable is used for long engineering and data builds; the choice is made at session start before session length is known.Source: §5.2 · p. 13

5.3 Holds cost money, not quality

• 42% of routable turns are held; two thirds of the $144k monthly gap to perfect labels comes from holds, while misroute correction chains account for the other third.The policy holds turns when classifier confidence is below threshold; a wrong move risks quality and rework, while a hold forgoes savings.Source: §5.3 · p. 13
• 63% complexity accuracy and one-level-off errors point to fuzzy taxonomy boundaries, while 12.5% of moved turns are false downgrades.The proposed remedy is sharper level definitions and examples tagged on enterprise prompts, with a threshold set on that data.Source: §5.3 · pp. 13–14

5.4 Pauses, the price sheet and tool use move the saving most

• Without pauses, savings reach 24%; with TraceLab’s more pause-heavy users, they fall to 7%, while a one-hour cache lifts savings to 17%.Under the stated alternative cache-price and write-premium assumptions, savings fall to 5%; without a Fable 5.1 licence they fall to 2.3%.Source: §5.4 · pp. 14–15

5.5 Strong models of other vendors cost 0.12 to 0.80 times Opus 5

• 0.12–0.80 times Opus 5 is the cost of strong other-vendor models on the same sessions when cache reads are priced; complex-work routing adds 10–55 points to savings.Without cache pricing, strong-model complex-work routing adds 3–24 points; these comparisons assume equal quality and token use.Source: §5.5 · pp. 15–16

6 Recommendations

• Two decisions have different horizons: put a router and control plane behind existing harnesses this quarter, then decide over the next one to two years whether to own the harness.Source: §6 · p. 16

6.1 Now: put a router behind today’s harnesses

• Put a gateway between every harness and model endpoint, preserve cache markers and beta headers, route at session start, and export aggregate token telemetry.Build the routing taxonomy from each function’s prompts; set a fleet default for subagents now and route them per launch when supported.Source: §6.1 · p. 16
• Keep the shipped bundle short: always-loaded prompts, tools, skills, and agent descriptions are reread on every tool step, and one measured 55k-token prefix cost three to four times the lean default.Deferred loading of another 104k tokens of tool schemas avoided still higher cost; audit per-category and per-tool accounting.Source: §6.1 · pp. 17–18
• Set context defaults by role: prune or summarize on cold returns, but preserve reasoning within active engineering tasks and clear stale tool results in batches.The passage recommends testing answer quality with a pilot group before setting standards centrally.Source: §6.1 · pp. 18–19

6.2 Later: decide on purpose whether to own the harness

• An owned harness is appropriate for regulated workloads, enterprise-served models, or sufficient volume to justify a product team, but not as the default.Vendors ship weekly, while an open-source base shifts dependency to its maintainers; decide after the earlier control-plane work produces numbers.Source: §6.2 · p. 19
• Keep model choice open through an enterprise endpoint, and evaluate vendors on cache-marker forwarding, endpoint access, subagent control, telemetry, retention, version pinning, and formats.The larger potential saving for complex work depends on which models the enterprise accepts and on harness endpoint support.Source: §6.2 · p. 19

Related work

• Routing studies include preference-based win prediction [33], confidence cascades [34], independent-query benchmarks [35] [55] [56], budgeted sequential formulations [57], and same-engine agent-serving simulations [58].The paper distinguishes its focus by pricing a vendor’s prompt cache.Source: §7 · p. 20
• Shipped routers use cache-aware rules, while independent Jev-based routers use safe-tier fallbacks or protect long conversations.OpenAI’s GPT-5 router drew user backlash over lost control, motivating a design that shows the model and allows a pin.Source: §7 · p. 20
• The nearest academic study evaluates agentic caching without model switching [65]; Codex cache measurements examine model and effort changes.This paper takes the cache as given and asks which model should own it.Source: §7 · p. 20
• The skill argument draws on expertise analysis, the METR trial [27], and DORA, while context-file ablations [30] [31] leave the unskilled-user case open.The paper also says EU AI Act and co-determination constraints shaped its omission of a productivity axis.Source: §7 · p. 20

8 Limitations

• The estimates rely on an emulation, two public corpora, and one vendor family’s list prices; lock-in repricing assumes equal quality and token use across models.There are no contract discounts; differing model discounts or price changes can move percentages, and under a uniform 0.1× cache-read rule the crossover disappears and savings fall to 5%.Source: §8 · pp. 21–22
• The 46 conversations and 109 turns are synthetic and author-tagged; measured behaviour perturbs them but adds no new tasks.Job types without a public corpus borrow developer behaviour on turns without tool work; the authors call for real, independently tagged prompts where available.Source: §8 · p. 21
• The routing taxonomy is unfinished: complexity levels are least crisp, there are no ablations or inter-rater agreement results, and team accuracy requires independent prompt tagging.Results span current labels to a perfect-label ceiling; the classifier’s accuracy for a team is known only after two independent raters tag its prompts and report agreement.Source: §8 · p. 21
• Only 6.0% to 12.6% of an all-Opus bill is the crossover on repriced real sessions, versus 12.2% in the emulation.Engineering scenarios under-represent turns with fewer than seven requests and have higher cache re-reads; TraceLab-shaped sessions would bring the modeled 14% saving to about 8%.Source: §8 · p. 21
• At $13 per engineer-day, the modeled bill is $10.0M a year and savings are $1.3M to $2.1M; shorter sessions also lower the saving percentage.The base case spends $40 per active day, between TraceLab heavy users’ $15 median and $58 mean; the blended rate changes when engineering’s share of the bill falls.Source: §8 · p. 21
• For small organisations, savings may not cover the engineering team needed to build a router, though standard routers requiring no in-house engineering may remove that threshold.Source: §8 · p. 21
• The non-engineer sensitivity assumes that half or all users choose Fable 5.1, but that choice reflects the authors’ rollout experience, not measurement.Those cases yield savings of 15.3% and 16.6%, respectively.Source: §8 · p. 21
• The session evidence covers about 10,000 Claude Code sessions from two corpora of developers, which disagree on subagent use and pauses.They differ by a factor of four in subagent use per tool turn and 1.6 in pauses.Source: §8 · p. 21

9 Conclusion

• The price sheet sets rates, but the harness chooses which rate applies and how many tokens are bought, leaving much of the bill to vendor defaults when enterprises negotiate only prices.Source: §9 · p. 22
• At 10,000 seats, the router recovers 14 to 21% of model spend, or $3.3M to $5.0M a year.Jev labels prompts against a bring-your-own taxonomy; routing acts only where no running conversation must rebuild its cache.Source: §9 · p. 22
• Appendix A specifies the taxonomy and policy and Appendix B the cost model and constants, but the simulator, scenario catalogue, and cached classifier labels are not released.Source: §9 · p. 22
• The reported spend, prices, populations, and savings come from public list prices and session corpora, the authors’ sessions, and an emulated enterprise.They do not represent Accenture’s spend, prices, contract terms, or internal figures.Source: §9 · p. 22

A The routing taxonomy

• Labels, descriptions, and examples in the bring-your-own taxonomy are editable configuration; only four context values are fixed because execution modes depend on their meaning.The case-study default has four axes and two modifiers.Source: A The routing taxonomy · p. 28
• The action axis has 13 domain-neutral verbs, including lookup, explain, build, diagnose, operate, steer, and converse.Steer is a short control message within an ongoing task and inherits that task’s tier.Source: A The routing taxonomy · p. 28
• The 12 domain labels drive policy floors and expected session shape, not difficulty.Source: A The routing taxonomy · p. 28
• Complexity runs from trivial through routine and complex to expert, mapping to the fast, balanced, strong, and frontier capability tiers.The levels range from a one-step task to novel, ambiguous, high-stakes, or long-horizon work where errors compound.Source: A The routing taxonomy · p. 28
• Context distinguishes none, reconstructible artifact, conversation history, and external information; the history distinction matters to routing decisions.Source: A The routing taxonomy · p. 28
• Stakes may raise the tier but never above the user’s model, while short, medium, or long output shape informs cost estimates.Source: A The routing taxonomy · p. 28
• The policy acts only when both the probability that the candidate tier suffices and the probability of no history dependence clear a threshold, set to 0.8 in the case study.Steer, task-opening actions, and turns in an open task stay on the home model; legal, finance, and HR have a balanced-tier floor, as do agentic edits.Source: A The routing taxonomy · p. 28

B Cost model and constants

• Each API request is charged for uncached input, cache reads, cache writes, and output, while the simulator tracks per-model cache lifetime and compaction.It also charges side-lane ledger construction, cache writing, and the home refresh; compaction at 85% of the window replaces history with a 20k-token summary.Source: B Cost model and constants · p. 29
• Misrouted turns are redone on the home model when their required tier or history exceeds the destination model; a too-weak session start pays its first turn twice.The naive-switch baseline uses the same misroute charging rule.Source: B Cost model and constants · p. 29
• Subagent costs are modeled from tool steps, prompt and brief size, tool-result tokens, output, and final-report length, with light, typical, and heavy profiles.The profiles use 5, 12, and 60 steps, respectively, with P = 16,200.Source: B Cost model and constants · p. 29
• Each scenario is expanded into 200 weighted variants; corrections and pauses follow SWE-chat distributions, while tool-work turns launch subagents with probability 0.039.Variants use a fixed seed and separate random streams so each sensitivity changes one behaviour at a time; correction chains and measured tool calls are also modeled.Source: B Cost model and constants · p. 29
• The case study uses 10,000 seats across 14 segments and 301,000 sessions monthly, with engineering accounting for 86% of spend.It applies 21 September 2026 list prices, a 5-minute cache lifetime or a 1-hour sensitivity, and specified prompt, ledger, classifier, and token-counting constants.Source: B Cost model and constants · p. 29
• Authors’ scripts produce the numbers, figures, and tables from the catalogue, cached classifier labels, and two public corpora, with an audit script recomputing each printed number.The code and catalogue are not released; the appendices specify the taxonomy, policy, and cost model so enterprises can rebuild the method with their own inputs.Source: B Cost model and constants · p. 29

• Twenty agentic coding harnesses are graded on the properties an enterprise inherits.Source: C The harness landscape · p. 30

• Figure 11 compares the cost of one subagent across models and activity profiles, noting the top-to-mid-tier gap is largest for short subagents and narrows for very long ones.The setup uses a 15k-token system prompt and 1.2k-token brief; subagents write context once and read it repeatedly.Source: Figure 11 · p. 30
• Figure 12 plots saving, false-downgrade rate, and hold rate against the confidence threshold for both classifier decisions.The decisions concern the tier a turn needs and its dependence on conversation history.Source: Figure 12 · p. 30
• Table 8 compares twenty harnesses by licence, model coupling, own-endpoint support, subagent model, Skills/MCP, and telemetry.Source: Table 8
• Figure 13 shows annual saving by model spend per engineer-day and seat count, from current labels to perfect labels.The outlined cell marks the case study; engineering spend scales with spend level, while other spend and segment percentages remain emulated.Source: Figure 13 · p. 32