Welcome to the agent platform research briefing for Thursday, August 7th, 2026.

**Google DeepMind leadership earthquake** — The biggest AI organizational shakeup of 2026 just hit. Jeff Dean, Google employee number 30 and one of the architects of modern machine learning, has left the company after 27 years to found Discovery Loop — a new startup aiming to automate machine learning, science, and engineering. He's taking three key people with him: senior fellow Sanjay Ghemawat, DeepMind VP Oriol Vinyals, and Google Brain co-founder Quoc Le. Same day, Demis Hassabis stepped down as CEO of Google DeepMind to become its Chair and Chief Scientist of Alphabet. Koray Kavukcuoglu, previously CTO of DeepMind, now runs the unit as Senior VP reporting directly to Sundar Pichai. Google is consolidating all AI leadership at its Mountain View headquarters in response to accelerating competition from Anthropic and OpenAI. Alphabet stock dropped roughly 4-5% on the news. For context, this comes months after Pichai admitted Google was "a bit behind" on agentic coding, Gemini 3.5 Pro was delayed, and Gemini 4 pretraining is now underway as Google's largest training run ever. The question now: can the new leadership structure deliver Gemini 4 before the competition pulls further ahead?

**Claude Code goes self-hosted with cross-session messaging** — Anthropic launched a major enterprise feature set for Claude Code this week. Sessions can now run on infrastructure your team controls — your own compute, your own sandbox, your own security posture. Cross-session SendMessage lets Claude Code instances message each other across machines on macOS and Linux, with ListAgents for discovery. New plugin and sandbox controls give admins finer-grained policy enforcement. Anthropic also improved Remote Control availability messaging, paste handling, and session reliability. This is a significant move for enterprise adoption: organizations that couldn't justify sending code to Anthropic's infrastructure now have a path. The cross-session capability effectively creates a distributed multi-agent mesh — one Claude Code can kick off work on another machine and coordinate results.

**Azure API Management launches dedicated AI Gateway tier** — Microsoft released a new AI Gateway tier of Azure API Management in public preview, with a control plane explicitly built around models, MCP servers, and tools rather than traditional APIs. It fronts Foundry-hosted models from OpenAI, Anthropic, and Mistral alongside AWS Bedrock, Google Vertex AI, and direct OpenAI connections — all behind one endpoint. This signals that MCP has graduated from experimental protocol to enterprise infrastructure primitive. Microsoft is building the plumbing that lets organizations route, govern, and audit AI tool calls across providers through a single control plane. Also notable: OpenAI Codex released support for the opt-in MCP 2026-07-28 stateless protocol just hours ago, adding paginated discovery, multi-round requests, and non-blocking server startup. Cloudflare published an in-depth blog on running stateless MCP on Workers with zero infrastructure overhead. The stateless spec ecosystem is now operational across the major platforms.

**SpaceX Starship Flight 14 targeting late August with first tower catch attempt** — SpaceX is planning Flight 14 for late August 2026, and it's going to be the most ambitious Starship test yet. Using new V3 Block 3 vehicles — Booster 21 and Ship 41 — it will be Starship's first operational orbital flight, deploying Starlink V3 satellites. But the headline event: pending regulatory approval, SpaceX will attempt to catch the 171-foot Ship upper stage with the tower's chopstick arms at Starbase in Texas. They've been catching the Super Heavy booster since Flight 5; catching the upper stage is an entirely new challenge. Musk confirmed the attempt is tentatively scheduled for end of month. If it works, it proves full reusability of both stages in a single mission — the architecture that makes Mars economics possible.

That's the briefing for today.