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

What it claims

The paper puts hard numbers on a claim the summary only states qualitatively: for coding agents, the cache, not the list price, decides cost, and the harness controls the cache. It builds the cost model, derives a payback rule for switching models mid-session, and shows via ~10,000 repriced real sessions and a 10,000-seat emulation that a cache-safe router recovers 14–21% of model spend ($3.3M–$5.0M/yr off a $23.7M baseline, ~$40/engineer/active day). Authors are from Accenture Responsible AI and Harvard Extension.

How it works

The unit priced is the turn, not the request: one user turn is k+1 API requests, each re-sending the whole conversation over a prompt cache bound to one model. Moving a turn to another model discards that cache and pays a cold write. So the router only switches where no live conversation must rebuild its cache — session start, side lanes, subagent launch — acting only on Jev's confident calibrated labels, and never downgrading a correction.

The mid-task payback threshold is n* = (Wt + Wh − Rt − Rh)/(Rh − Rt) requests (R = cache-read price, W = write price). Concrete: Fable 5.1→Sonnet 5 pays back only after 291 requests (46 without return); Opus 5→Sonnet 5 after 27 (8 without).

Evidence

• List price is the wrong metric. Anthropic's 21 Sep 2026 sheet (per M input/output tokens): Fable 5.1 $10/$50, Opus 5 $5/$25, Sonnet 5 $2/$10, Haiku 4.5 $1/$5. But Fable's cache reads are 0.025× input vs 0.1× for others, so the top model costs less than the next tier on 51–80% of real sessions over $100. A one-line change in a long session drops $0.88→$0.36.
• Pauses matter as much as the classifier. The API cache lasts 5 min; 17–27% of turns follow a longer pause and re-pay for the whole conversation. Savings range 24% (no such pauses) to 7% (heaviest pausers); a one-hour cache adds ~3 points.
• Corrections are over a third of the bill — hence never downgrading them.
• Lock-in price: other vendors' strong models cost 0.12–0.80× Opus 5 on the same sessions, but Claude Code (proprietary, Anthropic-only, non-Claude gateway routing explicitly unsupported) cannot reach them.
• Coupling matrix over 20 harnesses (8 Chinese-vendor, mostly MIT/Apache accepting any OpenAI/Anthropic endpoint); openness and model coupling are independent axes.

How to use it

The recommended control plane, runnable now without owning the harness: put a cache-safe gateway in front of every harness; choose models at session start and subagent launch; turn on the one-hour cache; report token use in aggregate only (EU AI Act flags per-user monitoring as high-risk); let each function define its own task taxonomy; ship one bundle of procedures (Agent Skills/SKILL.md) and connectors (MCP). Portable already: MCP, AGENTS.md, Agent Skills. Not portable: the control plane itself — model policy, telemetry, hooks, subagent behaviour, billing.

Limits

The 10,000-seat figure is an emulation — a tagged catalogue perturbed with behaviour measured in public session datasets, with a live classifier and misroutes charged as correction chains — not enterprise telemetry. All dollar figures are pinned to Anthropic's 21 Sep 2026 list prices. The fetched text cuts off mid-Section 4.2, so the full router spec, emulation details (Section 5) and control-plane architecture (Section 6) aren't visible here.