AI Builders Digest

Agents are already the majority of inference and heading toward nearly all of it

Levie argues the token mix has already flipped: agents, not humans typing into chat boxes, consume most inference today, and within a year or two that goes to nearly all of it. The work he expects them to absorb is unglamorous and enormous, reading every code change for security, processing data inside workflows, doing the bulk of recruiting and prospecting research, reviewing every event stream and log from every system in the world. His evidence is personal rather than projected: in the past week alone he stood up multiple workflows that were not technically possible a month ago.

• #agents
• #products

Claude Projects coordinates your sessions so you can fire off tasks and walk away

The new Projects experience puts Claude in charge of coordinating every session rather than making you manage them. Wu describes operating at a higher level of abstraction: batch out tasks, move on, and ask for an aggregated status update whenever you want, because the project holds context on everything you have going. It also carries long-lived memory that evolves with use. Rolling out over the coming weeks.

• #agents
• #products

Projects is one persistent agent per project that spins off subagents

The architecture under Projects is a single agent per project that owns memory and delegates tasks to subagents, brought over from Claude Tag. You can tell it to work proactively or on a schedule. Thariq's pitch is comparative rather than technical: it simply feels nicer than juggling a pile of separate sessions.

• #agents
• #products

Project threads run in the cloud and share memory, local access coming soon

Each thread in a project both adds to and draws from shared memory, so Claude retains things like the release slipping to Friday or who to check with before touching the billing service. A project library holds the files you add alongside the ones Claude creates. The tradeoff right now: threads execute in the cloud and keep running while your machine is off, but cannot reach local files, tools, or your internal network yet. Existing Pro and Max projects keep working unchanged until the rollout reaches them.

• #agents
• #products

Vercel went from 1 billion deployments in 10 years to 2.4 billion in 10 months

Rauch expects more software to be produced next year than in all of computing history, and he has a leading indicator: Vercel took a decade to reach its first billion deployments, then added 1.4 billion more in the last ten months. Much of that new volume is tiny, personal, even disposable software, an HTML file with some JavaScript, a report, a pricing calculator, a slide deck. Deploy time for an artifact, from upload through domain assignment to global propagation, is now one second including CDN, firewall, immutable deployments, observability, and rollbacks. There is also a new `vercel --turbo --prod` that grabs the fastest available build machine for panic hotfixes.

• #products
• #agents

A one-person self-driving site hit 6k daily impressions with zero backlinks

After a year of trying self-driving company products and finding none that cleared the bar, Kothari built his own: nosugarforkids, a catalog of healthy kids' snacks with a nutrient-based tier list and a chat interface for school lunch and snack box recommendations. A Claude agent wakes up once a day to check for new products, prune dead ones, pull performance from DataForSEO and Search Console, write and edit new pieces, and now pitch publications for backlinks, all while trying not to introduce slop. With no backlinks and no social presence, it has grown organically to roughly 6,000 impressions and 60 clicks a day. There is a full MCP and WebMCP if you want to point your own agent at it.

• #agents
• #products

Google Labs launches a family agent that handles logistics for up to 5 members

The new CC is an AI agent aimed squarely at household coordination. It supports up to five members, sends a shared "Your Day Ahead" morning brief by email, autosyncs schedules and to-dos with a shared Google Calendar and Tasks, and takes delegated work in Google Chat like weekly meal plans and school supply lists. It will also fill out paperwork such as permission slips under your direction. The memory model is the interesting part: it tracks what applies to the whole family, like grocery lists and favorite restaurants, separately from what applies to one person, like dietary restrictions and timezones. Waitlist is open, US only, 18+.

• #agents
• #products

"I stopped managing sessions" — Projects changed how one of its builders codes

Cherny says Projects changed not just how he interacts with Claude but how he writes code. He no longer manages sessions at all; he sends thoughts as they occur, Claude splits them into threads, and the project accumulates a model of how he works. It is now where he does a large share of his coding.

• #agents
• #products

Memorable does agent memory with embeddings instead of spending more tokens

Tan flags Memorable's approach to memory as a genuinely new direction: optimize with embeddings rather than throwing more tokens at the context. He also pushed on alignment in a separate post, arguing the case for aligning to humankind rather than to any other objective.

• #agents
• #open-source

Interpretability, not policy, is the right path to alignment

Agarwal calls interpretability work the correct way to solve alignment and safety, and names Goodfire as the leading company doing it outside the frontier labs. His framing is not incremental: he calls the work generationally important.

• #policy
• #agents

Gemini's notebook app is underused on mobile, says the Google Labs VP

Woodward's pitch is narrow and specific: if you are not running Gemini's notebook product on your phone, you are missing the best version of it. He also points to a family use case his own household has adopted.

• #products