AI Builders Digest

Cursor shattered the assumption that dev tools top out at low-billion exits

The market size for agentic coding was far larger than almost anyone modeled, and the space was assumed saturated right when Cursor got going. The playbook worth copying: find the right product shape for agentic work, act as a neutral layer between models and workflows, post-train models only where it drives down cost or lifts performance, build infra specific to the workflow, and pair it with a category-aligned go-to-market. There was far more room to innovate between the user and the underlying model than people believed.

• #agents
• #products

A hundred passing evals tell you nothing about whether an agent generalizes

Basis builds agents that prepare entire tax returns end to end, thousands of inference steps deep, and the core lesson is that outcome-based evals are not enough. "Even if you got it right a 100 out of a 100 times, if a person is just getting it right because they're going to Wikipedia, the accounting firm wouldn't hire them, and so they shouldn't hire us either." The fix is behavior specs: markdown files, never shown to the agent, that an agent-judge grades trajectories against, so you reward the process and not just the answer. The sharpest jab lands on engineers who obsess over abstraction while their context is a mess, because at runtime the English affects performance and the code organization does not.

• #agents
• #evals
• #open-source

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

Apple developers can run fast local tasks on the on-device model and hand off to Claude when a request needs multi-step reasoning, code generation, web search, or code execution, streaming the response back into the same SwiftUI view. Because the framework returns typed Swift values from @Generable annotations, the Claude call receives clean structured input instead of raw user text. Available tomorrow on iOS 27, iPadOS 27, macOS 27, visionOS 27, and watchOS 27, authenticated with an Anthropic API key.

• #products
• #agents

3.7 Flash lands in the Gemini app, Pomelli turns product photos into video

Gemini 3.7 Flash is now live in the Gemini app. Pomelli, the Google Labs experiment aimed at small businesses, added the ability to turn product photoshoots into short videos and gifs, extending the same one-click polish that made it popular for still images.

• #products

Cheaper software will mean more software, not less, per the steam engine

More efficient steam engines increased coal demand, higher-level languages produced far more code, and cheaper spreadsheets produced far more financial analysis. AI plus software engineering should follow the same curve: far more software, aimed at problems that were previously not worth solving. Also argues Cursor's cultural impact is underrated, because "cursor for x" gave the industry a product pattern to escape the chatbot phase, and that when everyone can build with AI the differentiators collapse to product sense, domain knowledge, distribution, and execution.

• #agents
• #products

AI is not jagged, it is just AI-shaped

Calling AI "jagged" is like calling dogs jagged for being better or worse than humans at particular tasks. It measures a different kind of thing against a human baseline that was never the right frame. Separately, if you believe your colleagues are smart, their ideas came from somewhere real, and some of Linear's best shipped work has been fusions of several ideas or derivatives well downstream of the original concept. And there is no force in tech as destructive as the PM promo packet.

• #products
• #agents

X's anti-spam model never reads the post, so AI content mills walk right through

Reading X's open-source algorithm turns up TweetSpamBot, a behavioral model that analyzes up to 512 recent account actions for signals like posting bursts, quote-post behavior, timing, and dwell between actions, flagging patterns such as TWEET_CREATE_BURST, QUOTE_TWEET_SPAMMER, and CONTENT_AMPLIFIER. A high score can trigger an account challenge but does not appear to downrank the slop itself. The gap is that the model never looks at post content, so a mill can find viral posts and quote them ten-plus times a day with the same hook-numbers-backstory-takeaway template across unrelated topics and stay clean.

• #products
• #policy

The OpenClaw team moved to building OpenClaw with OpenClaw

Sharing agent sessions as URLs turns out to be the unlock, letting the team hand off in-progress agent work the way they would a link to a doc. Also added an instruction to the shared AGENTS.md requiring any PR that changes UI state to upload a video of it, which pushes visual review into the diff instead of leaving it to reviewers to imagine.

• #agents
• #open-source

ChatGPT can now book restaurant reservations

Quick restaurant reservations now work directly in ChatGPT, shipped alongside a batch of other updates. Open question posed to Codex users: what hard problem did it solve for you this week, and where do you actually learn from others about pushing the frontier?

• #products
• #agents

With Fable 5, "take all recommendations" became a safe answer to one-way doors

Using GStack before and after Fable 5, the surprising shift is that many one-way-door questions Claude Code raises can now just be answered with "take all recommendations" and you end up happy with the result. Separately, on housing: the opponents are effectively assembling a comprehensive list of California laws that YIMBYs need repealed.

• #agents
• #policy

You can be an AI-native rocketship without permanent fundraising

Rejects the premise that fast-growing AI companies must live in a constant death match for customers, sacrificing gross margin and staying perpetually in market for capital. That path exists, but so does the other one, and the rules for building the second kind of company are very different.

• #funding
• #products

Vercel claims the fastest AI Gateway infrastructure in the world

Straight performance claim on Vercel's AI Gateway, positioning routing latency as the competitive axis for the layer sitting between apps and model providers.

• #products

Build personal iOS apps through TestFlight, no App Store required

TestFlight is an underrated distribution path for apps you never intend to publish. If the app is just for you or a handful of people, you get a real installed iOS app without touching App Store review.

• #products

AI Engineer NYC wave 1 CFP acceptances close today

Last call for wave 1 of the AI Engineer NYC CFP, with acceptances being finalized today and special requirements attached to the mainstage finance keynotes. Last year's NYC event was the most successful summit the series has run. On the Databricks fundraise: the M in the $188B series M stands for "we are going to kill so many meetings."

• #funding
• #products

AI compresses the workday into pure decisions, and that is exhausting

Before AI a workday was one decision followed by a long tail of process, repeated until 10pm. With AI it is decision after decision after decision, and by 3pm the brain is empty. The processing was doing more restorative work than anyone accounted for.

• #agents
• #products