06-reference

every google io agents anthropic acquires figma vibe check

2026-05-20·reference·source: Every·by Jack Cheng
google-ioanthropicstainlessmcpagentsfigmaclaude-stack

Every — Google I/O: Agents, Agents, Agents

Why this is in the vault

Three convergent threads in one issue, all live in the active RDCO investing/strategic frame:

  1. Anthropic confirmed acquiring Stainless for a reported ~$300M (per The Information). Stainless is the API-tooling layer used by OpenAI and Google themselves — Anthropic just bought the picks-and-shovels vendor its rivals depended on. This is signal #N in the Claude-stack cluster (after Karpathy joining, State of AI Dev 2026 Claude dominance, sandboxes/MCP tunnels, and today's Stratechery + Innermost Loop pieces).
  2. Google I/O collapses agents into the default Search box — same event Stratechery dissected today, but Every's framing splits into "collaborate-with" vs "delegate-to" buckets, which is a cleaner mental model than Thompson's world-models framing.
  3. Figma shipped an in-canvas agent — paywalled body teases evaluation, relevant to RDCO design-stack decisions surfaced in today's Hallmark eval.

Sponsorship

Ghost (Postgres designed for agents). Sponsor block placed mid-Spotlight, before the Signal section. Pitch: spin up unlimited ephemeral databases for parallel agent experiments, MCP-native, free tier with 1TB. Disclosure clearly labeled as advertisement block. Not interfering with editorial — Stainless content stands on its own news value.

Also disclosed in the body: Dan Shipper (Every founder) is a small investor in Stainless. That disclosure means the Spotlight is not arms-length coverage of the acquisition — Every has financial upside. Read the framing accordingly.

Issue contents

Spotlight — Alex Rattray, Stainless CEO and MCP whisperer. Re-up of the October 2025 AI & I episode with Rattray, surfaced because Anthropic just bought the company. Episode covers MCP-server design principles: keep tool count small, name tools precisely, generate tightly-defined outputs. Anthropic's official acquisition note framed around "extending Claude's ability to connect to data and tools."

Signal — Google goes all-in on agents. Gemini 3.5 Flash launched as the engine (~4x faster, ~half the cost vs comparable LLMs per Every's repeat of Google's framing). Splits launches into:

Inside Google I/O — Anyone can cook (paywalled). Tease about whether Gemini 3.5 Flash crosses the "good enough" threshold for everyday tasks. Behind paywall.

Mini-Vibe-Check on Figma's in-canvas agent (paywalled). Tease: what it gets right + still wrong, blank-page problem evaluation.

Mapping against Ray Data Co

Claude-stack thesis — STRONGEST signal of the week. The Stainless acquisition is not a "nice-to-have" tooling deal — Anthropic acquired the API-design layer that OpenAI and Google were paying to use. Combined with the cluster:

That's 9+ signals in ~72 hours all pointing the same direction: Anthropic is consolidating the developer/agent surface layer while Google flanks with consumer-agent ubiquity. For RDCO's investing thesis on the Claude stack, this is a corroborating data point worth flagging for the cluster pulse. NOT yet a paper-trade trigger — Anthropic is private, and the public-market proxy chain (Amazon, hyperscaler infra) is already loaded; this just thickens the case.

Google framing comparison vs Stratechery. Every's "collaborate-with vs delegate-to" axis is cleaner than Stratechery's "world-models" framing for product-strategy purposes. Worth noting Thompson's piece leans into Gemini-as-platform-bet whereas Every leans into Gemini-as-product-feature-storm. Both observations true; the split helps when reasoning about Google's actual strategic priority order.

Figma agent — DEFER until paywall content accessible. RDCO design-stack question stays open (Hallmark eval surfaced overlap concerns earlier today). Every's vibe-check format is one of the more trustworthy evaluation surfaces in the AI-press economy. Worth paying attention to the unlock or finding a vibe-check summary elsewhere this week.

Stainless playbook for our own work. Rattray's three MCP-server design principles (small tool count, precise names, tightly-defined outputs) are directly applicable to any RDCO-built MCP integrations. Re-listening the October episode would be a reasonable use of 1 hour this week given the principles now have a $300M endorsement.

Related