AI Builders
Digest

Y Combinator President & CEO Garry Tan makes a bold prediction: systems of record will need to become AI harnesses or face replacement by agents. This points directly at the existential pressure traditional software infrastructure faces in the agent era.

OpenAI Codex & ChatGPT lead Thibault Sottiaux states that 2026 is the year companies start seriously caring about model efficiency and reliability as it becomes critical infrastructure. He also updates that the reset has been propagated to accounts with fixes for recently reported usage issues, so users should feel a positive difference.

Vercel CEO Guillermo Rauch observes that OpenAI Sol price reductions and discounts on the Vercel AI Gateway have made Sol their fastest-growing frontier model. This shows demand for intelligence is highly elastic: as inference costs fall, usage grows rapidly. Teams not using a gateway are missing out on this price volatility that lowers operating costs and increases margins. The router space has heated up because gateways are inevitable.

Meta AI Senior Director Madhu Guru shares part 7 of building great evals: the Goldilocks principle. Evals should measure at the level of the various jobs to be done, not just the final answer. For a financial analysis agent, intermediate stages—understanding the client, gathering evidence, analyzing data, making a recommendation—each deserve their own eval. When the final recommendation is wrong, a well-designed set reveals exactly where the drop occurred (e.g., data analysis at 70%), so teams know where to dig and whether to break steps down further. Not too granular, not too coarse—just right to diagnose and act.

Peter Yang highlights Shreya’s distinction between two types of AI evals. Top-down: what you would come up with in a vacuum given only the task description—Claude does a very good job helping with these. Bottom-up: the gut feedback that emerges after looking at lots of sample outputs and externalizing it into evals—“Claude is very, very bad at coming up with bottom-up evals. That’s all you.” Full discussion is available in the linked episode.

Vercel CEO Guillermo Rauch outlines the philosophy for extending fx: open protocols—MCP (modelcontextprotocol.io), Skills (agentskills.io), Plugins (agent-plugins.org)—and the best one, Unix. Small programs that do one thing well and compose by calling other programs; libfx that enables embeddability into more complex programs so you can build your own CLI, background agent, or software factory, local or cloud.

Claude Blog: Building intelligent apps for Apple platforms with Claude in the Foundation Models framework. Anthropic is releasing Foundation Models framework support for Claude via a new Swift package. Developers can use Apple’s on-device models for fast local tasks like summarization or extraction, then hand off typed Swift values to Claude for multi-step reasoning, code generation, web search, and code execution. Responses stream back into the same view. Available on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27—add the package, sign in with an Anthropic API key, and the package handles streaming, tool calls, and structured responses.

Peter Steinberger (working on OpenClaw and with OpenAI) notes that while CLI is nice, having UI visualizations and your team where you work is nicer. He also added the rotation USB protocol to his claw project and instructed it to look around with a 360 webcam.