Radar / Ideas / HarnessLedger: the calibrated decision…

HarnessLedger: the calibrated decision layer that makes agent fleets 50-100x cheaper

daily ideamoderateJEV confidence 0.54JEV monetization 0.582026-10-21

OutcomeCut an agent fleet's token spend 50-100x with no drop in task success.

The problem

Agents are chatty: a single user request triggers 4-12 model calls, each resending a growing context, and real bills land 8-15x above single-prompt estimates. Teams respond with hand-tuned routers and dashboards, but the routing itself stays artisanal — every team rebuilds the same two-tier escalation logic, and nobody measures the one metric that matters: cost per successful task.

The idea

A drop-in middleware that sits between an agent harness and the model APIs. Every per-step decision (tool choice, classification, routing, whether to escalate) goes through a two-tier router: cheap decision models (Mouse-1-class) handle the routine calls, Nemotron-class rerankers prune the evidence before the frontier model ever sees it, and Jev-style typed judgments gate the ambiguous ones with calibrated probabilities. Every decision is logged with its cost and outcome, so teams see cost per successful task — not cost per token — and the router's calibration improves per harness. Pricing is a share of the spend it eliminates.

Why now

The building blocks only just arrived together: Mouse-1 (Oct 21) proves a purpose-built decision model can match Opus 5.5 harness output at ~1/190th the cost; Jev gives the typed judgment layer for the calls too ambiguous to route cheaply; and NVIDIA's Nemotron 3.5 Rerank (Oct 21) shows evidence pruning alone cuts agent token burn ~14%. The agent-harness ecosystem (167 harnesses ranked by WeSearch) is now big enough that a shared optimization layer beats every team rebuilding routing by hand.

What it combines

Mouse-1 (cheap decision model for agent harnesses) + Jev (typed noul/score/choice decision layer) + Nemotron 3.5 Rerank (evidence selection that cuts agent token burn ~14%). Any one alone is a cheaper model call; together they become a full harness-optimization layer: route the small decisions to the cheap model, prune the evidence the frontier model sees, and gate the ambiguous calls with typed judgments — all measured against cost per successful task.

MVP

Weekend MVP: a proxy middleware for ONE harness (LangChain) that routes small decisions to a cheap decision model, logs every call with cost and outcome, and reports cost per successful task vs the frontier-only baseline. Deliberately skip: multi-harness support, the per-harness LoRA calibration loop, the dashboard, and the revenue-share billing.

How it makes money

Budget owner: the VP Engineering / platform team that owns the AI infrastructure bill (they already pay for Braintrust at $249/mo and LangSmith seats). Pricing anchor: Microsoft Decision-1 at $0.042 per million input tokens (Product Hunt) for decision-model pricing, and Braintrust Pro at $249/mo (voiceflow.com) for the observability budget this replaces/augments. Willingness-to-pay: Braintrust's own cost-efficiency page frames the buying conversation around 'cost per resolved ticket' with measured strategies; the Medium cost piece documents finance teams being surprised when real agent bills land 8-15x above single-prompt estimates. Replacement budget (they already pay for observability + tokens) beats new-budget creation.

The moat

The largest incumbents in token economics are the model providers themselves — OpenAI, Anthropic, Google sell tokens, so a tool whose entire purpose is cutting token consumption 50-100x cannibalizes their revenue; they will not ship it. Microsoft's Decision-1 exists but is Foundry-locked, i.e. a lock-in play, not an open layer. An open, harness-agnostic, provider-neutral optimization layer is a different business model neither will adopt: it would commoditize their distribution.

Distribution

B2B: sell to platform teams running agent fleets. Wedge through harness-framework integrations (LangChain, CrewAI, OpenAI Agents SDK) and channel partnerships with observability tools (Langfuse, Braintrust) that already see the spend but don't cut it.

Why it wins

Braintrust and LangSmith observe spend but don't cut it — they sell visibility. Mouse-1 and Microsoft Decision-1 sell a single model API, leaving teams to rebuild routing, escalation and evidence pruning per harness. HarnessLedger is the missing middleware: decision routing + evidence pruning + typed gating in one drop-in layer, priced on the savings it creates, provider-neutral where Microsoft is Foundry-locked.

Risks

Biggest risk: decision-model quality varies per harness, and a bad route costs more (in retries) than it saves. The MVP de-risks this by starting with one harness and measuring cost per successful task against the frontier baseline before adding more.

Build it with

Mouse-1: decision model for agent harnessesThe cheap per-decision model the layer routes to — ~190x cheaper than Opus 5.5 for harness decisions.
Jev 1.14.0 — TypeSafe's judgment layer updateTyped decision layer: noul/score/choice judgments with calibrated probabilities for the escalate-or-not calls the router cannot make cheaply.
NVIDIA Nemotron 3.5 Retriever collection: Embed 8B/1B, Rerank 8B, Parse 631MEvidence pruning: the reranker cuts an agent's total inference tokens per task by ~14%, compounding the routing savings.

Repo to start from

harnessledger: drop-in decision/routing middleware for agent harnesses with cost-per-successful-task tracking and a share-of-savings pricing model.

Evidence

Get the week's best AI launches, plus 3 ideas worth building

One email every Saturday. Ranked by traction, not hype. Free.