Builder Radar

Cutting the agent bill is turning into its own layer

July brought three tools for spending less on tokens. August added four more, each working on the bill from a different side. Varion compacts history and prunes tool schemas before a request reaches the provider, returning a header so you can check the savings yourself. Vynaris routes each request to the smallest model its evals support and returns a receipt: model served, provider cost, routing fee. Weckr wraps your existing client to track cost and margin per end customer. Openspender goes furthest. It gives agents their own wallets, capped per request and per day, to pay for models and tools over the x402 and MPP payment protocols. Nobody provisions a key for each service. Two months in, this looks less like a feature of other products and more like a layer of the stack.

Agents are getting a memory and a package manager

screenpipe records selected screen, audio and meeting context locally and serves it to agents over MCP and a local API. Its GitHub repo passed 21,000 stars when we looked in August. Linkly AI indexes local documents, images, audio and video into a search engine agents can query. Two skill registries arrived as well. agentskills.codes listed more than 19,000 installable skills for Claude Code, Codex and Copilot in August, each with source-health scans. Another skills marketplace was submitted four times by the same sender. The shared idea: an agent should carry context between sessions and pick up new capabilities the way a developer installs a package.

KugelAudio trained its own voice models instead of reselling

Most European inference providers rent GPU capacity and serve open-weight models. It is a real business with a thin moat, and it describes almost every provider in this group. KugelAudio is the exception: it trained its own text-to-speech models in Berlin, published one under MIT, and competes on latency against ElevenLabs and Cartesia. One company is not a trend. It is the harder path, and nobody else in July took it.

Three ways to spend less on tokens, none of them switching provider

KV Cache Store precomputes the KV cache for long prompts so you skip prefill on repeated context. Tokenade sits between a coding agent and the model and rewrites requests before they are sent, billing on tokens saved. Parity Layer shadow-tests cheaper models against your baseline and routes to them once your confidence threshold is met. Three products, three mechanisms, all arriving the same month. Three is not a category yet, but a year ago the answer to a large bill was to find a cheaper provider, and none of these do that.

Someone is selling reseller-priced tokens as infrastructure

Tokenware, Teamorouter, Tokens Forge, and one relay that resubmitted repeatedly under rotating throwaway domains share a vocabulary: 'channel,' 'settlement,' 'official-channel pricing,' and headline rates far below what the model providers charge. That is not what OpenRouter or Portkey do; those send your call to the official provider API and mark it up. A gateway that undercuts the provider's own price by 80 to 90 percent has to be getting the tokens from somewhere else. It is worth tracking separately from the legitimate routing trend, because it is a different business.

The infra that shipped was about running agents, not building them

Away from the gateways, the accepted infrastructure was operational. Dunetrace watches AI agents in production, with failure detection, Slack alerts, and root-cause analysis. falsifylab exposes finance data as MCP tools an agent can call. CC Switch is an open-source desktop manager for AI coding tools that unifies providers, routing, and usage. Less 'here is a framework to build an agent,' more 'here is how you keep the one you built from falling over in production.'