AI Builders Digest

Claude plugs into Apple's Foundation Models framework via a new Swift package

Anthropic released a Swift package that lets Apple developers call Claude from inside Apple's Foundation Models framework, shipping tomorrow on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27. The split is deliberate: on-device models handle fast local work like summarization and extraction, then hand off to Claude when a request needs multi-step reasoning, code generation, web search, or code execution. Because @Generable annotations return typed Swift values, the Claude call arrives with clean structured input instead of raw user text. The package handles streaming, tool calls, and structured responses back into the same SwiftUI view, so it reads as one experience to the user.

• #products

Cheap inference is proving demand for intelligence is highly elastic

OpenAI Sol's price cuts and discounts on Vercel AI Gateway made it Vercel's fastest-growing frontier model, which Rauch reads as proof that demand for intelligence expands rapidly as cost falls. His second claim is sharper: if you are not routing through a gateway you are leaving the price volatility on the table, and that volatility is now a margin lever. No wonder the router space heated up, he says, gateways are inevitable. He also laid out how fx gets extended: open protocols (MCP, Skills, Plugins) plus the Unix model of small composable programs, with a libfx that lets you embed the thing into your own CLI, background agent, or software factory.

• #products
• #open-source

Score every stage of an agent, not just its final answer

The most common eval mistake is building a golden set of correct final outputs and checking whether the agent got there. For a financial analysis agent, the recommendation is the last of four jobs: understanding the client, gathering evidence, analyzing the data, then recommending. Each stage produces an intermediate output and each deserves its own eval, so a wrong recommendation resolves into a diagnosis (client understanding 92%, evidence extraction 92%, data analysis 70%, recommendation 75%) instead of a shrug. The Goldilocks rule: not too granular, not too coarse, exactly as granular as you need to act on the result.

• #evals
• #agents

Prediction: systems of record become AI harnesses or get replaced

Tan's call is that the incumbent systems of record have one move available to them. Either they turn themselves into harnesses that agents drive, or agents replace them outright.

• #agents
• #products

OpenAI propagates usage resets and lands fixes after limit complaints

The reset has gone out to accounts and fixes shipped for the usage problems flagged the day before, with more landing tomorrow and a promise to keep communicating. Users should feel a positive difference. Alongside the incident work, a broader bet: 2026 is the year companies start seriously caring about model efficiency and reliability, because the models have become critical infrastructure.

• #products

Claude writes good top-down evals and bad bottom-up ones

Evals come in two kinds, and models are only useful for one of them. Shreya Shankar's framing: top-down evals are what you would invent from the task description alone in a vacuum, and Claude helps a lot there. Bottom-up evals come from staring at piles of sample outputs and externalizing your gut reaction into a check. "Claude is very, very bad at coming up with bottom-up evals. That's all you."

• #evals

A CLI agent is fine, an agent with eyes and a UI is better

Steinberger added rotation USB protocol support so his agent can drive a 360 webcam and look around the room on its own. His related argument about interfaces: the CLI is nice, but visualizations plus having your team where you already work beats it.

• #agents
• #hardware