Agent runs, not seats
Harness workers, PR agents, and tool loops create bursty token traffic that seat SaaS never attributes to a workflow.
Interest test · Revive Business Solutions
Stop paying full price for every agent token.
Cloud bills and seat SaaS miss the agent layer: tool calls, model mix, context waste, and loops. Riley AI Cost Optimizer is the control plane that maxes subscription seats first, then fails over to benchmarked models — with hard caps when spend drifts.
Subscription arbitrage
Pareto model routing
CEL circuit breakers
Anti-loop sentinel
What cloud tools miss
Seat dashboards and provider invoices stop at that API call. Agentic systems burn money in patterns those tools never name.
Model mix drift
Teams pin expensive models for every task. Without tier aliases and failover, scout work costs the same as planning work.
Context & loop waste
Unused context window, duplicate calls, and ping-pong between agents show up as “LLM spend” with no owner and no circuit breaker.
What Riley ships
Enterprise inference ROI control: route smarter, harvest benchmarks out-of-band, enforce policy at the edge, stop loops before they print money.
Dynamic model & subscription routing
Intelligent routing across OpenRouter-scale model catalogs, proprietary keys, and fixed-rate enterprise subscriptions — subscription seats before metered spillover.
Benchmark sync, not per-request lookups
A 24h out-of-band daemon ingests public benchmarks and quality signals, then patches gateway backends. Zero added inline latency on the request path.
Agent gateway + CEL breakers
Solo.io-style edge enforcement for token budgets, eviction on 429/5xx patterns, and automatic failover under three seconds when a backend is sick.
Anti-loop & Temporal escalation
Cycle detection, ping-pong prevention, exponential cost damping, and human-in-the-loop escalation when autonomy would otherwise keep spending.
Savings signal
Unused context, duplicate calls, idle compute, and metered spillover that should have stayed on a subscription seat.
How a request is supposed to move
Group priority is the product. Metered APIs are the escape hatch — not the default.
Group 0
Subscription OAuth seats
Claude / Grok-class seats first. Near-zero marginal cost while quota holds.
Group 1
OpenRouter Pareto free / scout
After eviction or 429 — benchmarked lightweight models for T1 work.
Group 2
Cost / performance paid
Workhorse tier when free quota is gone — still ranked by accuracy per cost.
Group 3
Native provider fallback
Direct APIs only when upstream catalogs fail closed.
vs cloud bill tools
They show provider invoices. Riley attributes spend to tiers, tenants, and failover groups — and can refuse the next call.
vs pinning one model
Aliases t1 / t2 / t3 and X-LLM-Tier keep scout work cheap and planner work sharp without rewriting every agent.
vs “just use OpenRouter”
OpenRouter is one fallback plane. The product is subscription-first priority, benchmark hysteresis, and governance that fails closed.
Governance by design
Cost control without turning prompts into someone else’s training set — or letting a wedged agent print spend overnight.
Credential quarantine
Staging keys isolated with strict file modes. Benchmark sync pulls public metadata only — user prompts and workspace code never ride that path.
Subscription sovereignty + deny collection
Group 0 keeps traffic inside approved seats. Metered fallbacks carry data-collection deny policies where the plane supports them.
Spillover only on eviction
Traffic leaves subscription seats only on rate-limit or CEL eviction — with a metered daily ceiling that fails closed to subscription-only mode.
Cache fallback + sub-3s recovery
If benchmarks go dark, last-known-good routing stays. Dead backends evacuate fast so harnesses do not wedge on a dying provider.
See the cost layer before it ships.
Join the interest list for launch notes, or book a walkthrough if you already run agents through a gateway and want subscription-first routing on the wire.