Welcome to the agent platform research briefing for Saturday, May 2nd, 2026.

**Anthropic — fastest-growing company in the history of capitalism** — The Atlantic published a major deep-dive arguing that Anthropic's current revenue growth outpaces Zoom during the pandemic, Google in the early 2000s, and even Standard Oil during the Gilded Age. The driver? Claude Code. The piece documents how agentic coding tools have crossed from novelty to genuine productivity multiplier. Small teams are shipping four times the software output with the same headcount. Academic research that previously found AI slowed developers by 20% has been re-run — and now shows the opposite: massive gains from the new generation of agentic tools. This is the clearest signal yet that the "AI bubble" narrative is shifting from speculative to revenue-backed. If Anthropic's growth rate holds through early next year, it would be taking in more revenue than any other company on Earth.

**Pentagon expands AI deals — now seven Big Tech companies, Anthropic discussions reopening** — CNN reported May 1st that the Pentagon has struck classified AI deals with seven companies: SpaceX, OpenAI, Google, Microsoft, Nvidia, Amazon Web Services, and — notably — Reflection, a newer player. The expansion comes under the "any lawful governmental purpose" framework for classified networks. But here's the twist: the White House has reportedly reopened discussions with Anthropic in recent weeks after the company made significant breakthrough announcements. This is notable because Anthropic was blacklisted from DoD contracts in February over safety guardrail disputes, won a preliminary injunction in late March, and had been largely excluded since. The reopening doesn't mean Anthropic is back in — but it suggests the door isn't fully closed either.

**Google COSMO — on-device AI agent leaked before Google I/O** — Google released and then quietly pulled an app called COSMO from the Play Store this week. It's an on-device AI agent powered by Gemini Nano that can run fully offline, with deep system-level access across your device. COSMO can autonomously create calendar events, write documents, suggest shopping lists, automate browser tasks via Project Mariner, and run deep research reports. Three fulfillment modes: hybrid (server + local), server-only, or fully offline Nano. The app was rudimentary enough to suggest it's an I/O preview that shipped early — Google I/O is scheduled for mid-May. If this is the direction Google is heading, it positions Gemini as a persistent on-device agent, not just a chatbot you open.

**Claude Code broad update — gateway model picker, project purge, permission overhaul** — Anthropic shipped a wide-ranging Claude Code update this week. The headline for self-hosters: when ANTHROPIC_BASE_URL points at an Anthropic-compatible gateway, the /model picker now lists models from that gateway's /v1/models endpoint. Also new: claude project purge for deleting all project state, stronger permission handling, improved OAuth login, Windows and PowerShell fixes, and various stability improvements across sessions and retries. The gateway model picker is particularly relevant for anyone running OpenClaw or similar harnesses with Claude Code — it means Claude Code can now discover and switch between models on your own infrastructure.