AI Builders
Digest

The Takeaway: AI agents will need their own bank accounts because traditional payment rails cannot handle the flood of sub-dollar transactions in the agentic economy, and crypto is the natural fit.

Brian Armstrong, cofounder and CEO of Coinbase and cofounder of longevity company New Limit, lays out Coinbase’s three big bets: the Everything Exchange (stocks, commodities, crypto, derivatives, prediction markets all on one platform), stablecoin payments, and Agentic Finance (AIFi). Roughly 76% of agent ecommerce transactions are under 30 cents; credit-card minimum fees make traditional rails unusable for them. “We don’t want the AIs to be unbanked.” Coinbase now offers a single-prompt tool that gives any AI agent a self-custodial wallet using the X402 protocol they incubated and contributed to the Linux Foundation. Agents can pay for data, tool calls, or services without a human intervening at every paywall.

Internally Coinbase is building recursive self-improvement: every agent change to a service ingests that service’s “brain” (past incidents, financial controls, A/B tests, PR history); human corrections are written back so one-shot PR acceptance rates climb over time. Armstrong himself is already shipping features by spinning up parallel agents. At New Limit they use AI to search transcription-factor combinations for epigenetic reprogramming, have successfully reprogrammed at least one human cell type in humanized mouse models, and will start Phase 1 trials next year (first indication: alcoholic liver disease hepatocytes), with the long-term goal of restoring youthful function across liver, vasculature, immune cells and skin.

“Crypto was really, really good for humans, and it’s going to be essential for AI.”

Madhu Guru, Sr Director of AI at Meta (previously led Gemini and Veo at Google), diagnoses the three most common reasons enterprise AI programs fail: applying old-school product playbooks (central AI teams staffed by trusted lieutenants, incremental mindset instead of experimentation and invention), under-investing in evals, and building AI platforms “for” the business from the outside so tools never match real workflows and judgment.

What works instead: hire leaders who have actually shipped AI products (technical depth plus the ability to earn trust), make evals a first-class citizen, and embed your strongest AI builders inside the functions you want to transform—build with finance, with sales, with support.

Practical AI educator Peter Yang is bluntly skeptical of “software factories.” Outside of verification and testing, he does not believe AI is yet capable of self-improving a product or building a new feature end-to-end without a human in the loop. One wrong assumption overnight turns the entire run into wasted tokens. He asks for concrete examples of products or features that have been built this way with no human defining requirements or reviewing the work.

Replit CEO Amjad Masad announced that Replit has acquired a business built entirely on Replit and expects it to be the first of many. The company continues shipping new capabilities (Routines with budgets) and is seeing a wave of Astra-powered 3D experiments from the community.

Builder Zara Zhang argues the “one-person company” idea is overrated. AI lets one person do far more, yet building something new is a profoundly lonely experience. You need someone to brainstorm with, suffer with, and celebrate with. Without tying yourself to the mast with another person it is extremely easy to lose motivation.

Thibault Sottiaux (Codex & ChatGPT at OpenAI) reported a heavy Astra shipping week—Images 2.5, GPT-Live-1, Agents API, Data Agent, ChatGPT for Financial Services—and DevDay is still ahead. He also detailed a quality reset after user reports: some legacy skills were over-triggering, an opt-in context experiment that caused early stops was disabled (affecting roughly 4-5k users), and misconfigured engines were removed. Community examples accelerated the fixes. He also welcomed the Git AI team; their open-source tool helps developers see exactly where coding agents contribute, and OpenAI will keep it open source.

Thariq (Claude Code at Anthropic) launched plugin evals (claude plugin eval init) so developers can check whether their skills still work after new model releases. He also notes that pass/fail scores alone are nearly useless these days—many benchmark failures come from overly strict hidden tests, and sometimes the model’s answer is more sensible than the expected result.

Vercel CEO Guillermo Rauch highlighted that Tailscale’s model router runs on Vercel AI Gateway and declared “AI Gateways are the new CDNs”—going direct-to-origin is brittle and DIY is painful. Box CEO Aaron Levie announced the ability to mount Box into agent sandboxes so agents can read and write files on their own “computer” the same way humans do. Cursor designer Ryo Lu showcased long-lived agents now available for big, persistent ideas.

Every CEO Dan Shipper notes that public benchmark scores tell you almost nothing about how a model performs on real work. For three years the team has run long-form vibe checks; they have now built an internal platform so everyone can create personal benchmarks from their actual day-to-day tasks, combining qualitative feel with quantitative measurement.

FirstMark partner Matt Turck shared his full 9/11 story as a young founder whose office was on the 53rd floor of the North Tower: almost no one was in yet, the one colleague present escaped, and the team restored from backups and was operational by end of day. SPC partner Aditya Agarwal recalled being in a CMU lecture hall and reflecting on whether he would have had the courage shown by the passengers of United 93.

Y Combinator CEO Garry Tan argued that scoring 1600 should unlock a second, harder test to measure higher excellence rather than banning the SAT and turning everything into a lottery. FPV Ventures partner Nikunj Kothari observed that “success has many fathers, but failure is an orphan”—VCs increasingly fight for attribution on the rare hot deals, and emerging GPs especially need founder references to claim their real contributions. Peter Steinberger showed Astra playing Doom via CUA in a cloud session and submitted a Linux key-reliability patch to trycua.