AI Builders Digest

Parag Agrawal: human click data is a bug, and agent search needs its own economics

Parallel is building web search for agents on the bet that agents will search a thousand times more than humans ever did, and that the ranking signal Google was built on is the wrong one. "Our view at Parallel is that human click data is a bug, and agent doing work with search should rely on agent feedback, not human feedback." The company skipped the day-one full index by launching a search agent that crawls at query time, trading latency for coverage while the index grew, then attacked latency last and shipped a product that cuts a three second budget to 200 milliseconds. Agrawal is also trying to fix publisher payments with Shapley values, estimating a source's worth by how much extra compute it would take to recover the quality lost without it, and he puts meaningful payouts to a wide range of content owners twelve to twenty four months out. Parallel is now a search and grounding provider for Google Cloud's enterprise agent APIs, sitting alongside Google search as an option.

• #agents
• #search
• #products

Vercel ships Run SDK and takes Connect to GA, both aimed at agent code execution

Run SDK is a secure eval for dynamic Code Mode execution: when an agent writes code, run it in a lightweight QuickJS secure context instead of spinning up a full sandbox, which Rauch pitches as faster and more cost efficient. Vercel Connect also hit GA, giving you an MCP client you can query on behalf of an authenticated user after running something like `vercel connect create notion`. His framing for both: "The hardest problem in building agents is secure connectivity to services and data."

• #agents
• #products
• #developer-tools

Claude now has one memory shared across chat and Cowork

Hand Cowork a task and it starts from what Claude already knows from your chats, including the project you talked through, your manager's preferences, or last quarter's client. Everything remembered is stored as a list of topics in Settings that you can read, edit, or delete, and memory updates on its own as you chat or on an explicit "remember this." Topics some people consider sensitive, like health or religious beliefs, stay out of memory unless you turn them on. Memory is on by default for Free, Pro, and Max.

• #products
• #memory

OpenAI adds a team tier for ChatGPT and Codex with no 5 hour limits

Sottiaux says demand for this one was heavy: it works like the Pro $100 plan but is built for teams and small companies. It bundles all ChatGPT, ChatGPT Work, and Codex features, connectors for Google Workspace, Slack, GitHub, and Microsoft 365, SAML, SSO and MFA, centralized billing, usage analytics with spend controls, and no 5 hour limits.

• #products
• #pricing

Sam Altman: "we made a chip and it is fast"

That is the entire post. No specs, no vendor, no availability, but it is the first thing he has said on silicon in this form.

• #hardware

Most evals fail because teams freeze them while user behavior keeps moving

Part 9 of Madhu Guru's eval series argues evals need a roadmap, not a fixed artifact. Take a financial research agent: week one users want a 5 page earnings report summarized, three weeks later they want five reports compared into a growth story, two months later they hand over 15 filings and transcripts and expect an investment thesis, and eventually they want portfolio monitoring with alerts. Each stage needs different evals, moving from short to long context, single turn to multi turn, passage citations to line citations, simple QA to synthesis, reactive chat to proactive agent. His practical version: map the dimensions usage will evolve along, mine production traces for shifts, and build P0 evals for the next stage before users get there. If your evals are stuck in week one while users are in week three, it shows up in churn.

• #evals
• #agents

Levie: the gap between models and enterprise workflows is where the value sits

Riffing on a post about applied AI strategy at scale, Levie argues the world does not want raw models and agents, it wants outcomes, and the premium goes to whoever converts tokens into real world results inside specific industries. What that actually requires is unglamorous: understanding context, driving change management, a harness that routes across models, connections into the critical business systems of a vertical, the UX of putting users next to agents in a workflow, and domain specific evals. He frames it as a window that is open right now to build the defining companies in each enterprise domain.

• #agents
• #enterprise

Why the public hates datacenter buildout, from someone in the industry

Agarwal calls the backlash totally unsurprising, since AI today mostly helps knowledge workers and the highest paid segment of the country. He expects the turn when AI starts finding cures for diseases that affect everyone, and says there is light at the end of that tunnel, which is also why some of the brightest people are moving into it. His self criticism of the industry: instead of painting a positive version of the future, we spent the time explaining why everyone should be afraid.

• #policy
• #hardware

Peter Yang open-sources /fuck-cancer, a skill for patients and caregivers

The skill builds and maintains a single source of truth brief from your documents and context, with five sections: patient and care team info for reference during calls, exactly three next actions, what we know split into confirmed facts versus open questions, medical terms in plain English, and a care log of recent updates and decisions. When it needs research it pulls from the National Cancer Institute and the ClinicalTrials.gov API. It runs with ChatGPT, Codex, or Claude Code, and saves either a local Markdown file or a shared Google Doc so the whole family works from the same information. It came out of what patients and caregivers told him after he shared his mom's story, including that the volume of doctors, documents, and insurance paperwork gets overwhelming fast.

• #open-source
• #products
• #health

Warning: Codex locked-use capabilities are locking people out of the macOS keychain

Swyx says it has locked him out twice this week, and points to Apple developer forums acknowledging it as a known bug in unstable Mac features that Codex currently relies on. His advice is blunt: avoid it for now. He would rather do everything in the cloud, but says the cloud is not there yet.

• #developer-tools
• #bugs

Google Labs opens a waitlist for Play with Putty, multiplayer vibe coding

Putty is a collaborative tool for building websites and small tools together in real time, framed as vibe coding going multiplayer. It is waitlist only, US only, 18+.

• #products
• #developer-tools

An El Niño monitor built solo with Codex and shipped on Railway

Kothari launched elneenyo.com, pulling real time updates and news straight from government sources, with per region impact and cost estimates, historical records, and a glossary and FAQ for the different readings. He built it with ChatGPT Codex on Railway, with design polish from Emil Kowalski's skills and generative loaders from Kasturi, after being inspired by an Odd Lots episode.

• #products
• #agents