Three verified moves this week, and each rewards a different instinct. One is a standard that finally stops moving — and quietly promises to keep standing still. One is the most safety-forward lab in AI putting the open-weights fight in writing, in a direction that helps builders. One is a seed round that says the hard problem in robotics isn't the brain. Every item below is dated and sourced, and each carries the single line that changes what you do next.
1. MCP ships final today — and the real news is that it promised to stop breaking you#
The Model Context Protocol 2026-07-28 specification locks final today, and it is the largest revision since the protocol launched. The headline you already knew: the core goes stateless. The Mcp-Session-Id header and the initialize/initialized handshake are gone at the protocol layer, so any request can be served by any instance — a remote MCP server can now sit behind a plain round-robin load balancer instead of sticky sessions and a shared session store.
Two changes underneath the headline matter as much. Tasks graduate from experimental to core: a server can answer a tools/call with a task handle instead of blocking, and the client drives the work with tasks/get, tasks/update, and tasks/cancel. The old blocking tasks/result and the tasks/list method are gone — without sessions, scoping a task list to the right caller was unsafe, so the spec chose the safer design. That gives you long-running agent work as a first-class, stateless pattern. And authorization is hardened across the board.
But the piece a founder should actually circle is the new feature lifecycle policy. Every feature now has an Active → Deprecated → Removed lifecycle with a guaranteed minimum of 12 months between deprecation and the earliest possible removal. Roots, Sampling, and Logging enter documented removal windows rather than simply vanishing.
What it means: For two years, building on MCP meant absorbing churn. The stability policy ends that. MCP is now a dependency you can plan around — you get a full year of notice before anything you depend on can be removed, which for a bootstrapped team is worth more than any single feature. Upgrade to the stable SDKs, then schedule your migrations off deprecated features inside their windows instead of scrambling when a spec drops. We mapped every feature to its lifecycle state — what's safe to build on today and what's already on the clock — in the MCP feature lifecycle map. The migration step-by-step is in how to migrate your MCP client to the 2026-07-28 stateless spec, and the async pattern is in how to run a long MCP tool call as a task.
2. Anthropic drew the open-weights line — and founders are on the safe side of it#
On July 27, Anthropic CEO Dario Amodei published Our position on open-weights models, one day after Kimi K3's weights went fully open. The core statement: Anthropic "has not and is not advocating for a ban on open-weights models as a category," and non-dangerous open models are "a public good." His three actual asks — tighter chip export controls, an industrial-scale distillation crackdown, and mandatory safety testing for frontier-capability models (with smaller startup and academic models exempted) — all aim at the supply chain and the frontier.
What it means: If you were holding off on an open-weights backend because a ban felt imminent, that reason is gone. The pressure is pointed past the application layer where a small team operates. We decoded each ask against a founder's actual stack — chips, distillation, and testing, one at a time — in Anthropic just drew the open-weights line, and founders running Kimi K3 are on the safe side of it.
3. Enigma's $71M seed bets the robot bottleneck is the interface, not the brain#
Enigma, an Israeli physical-AI startup founded under a year ago by Jonathan Jacobi and Gal Niv, emerged from stealth on July 27 with a $71M seed led by Index Ventures and Ribbit Capital. The interesting part is the thesis and the cap table. Enigma's bet is that the bottleneck to robots isn't intelligence — it's usability, the interface between a human and a machine. To prove it, the company put 100+ of its AI-powered robots online at robots.online, where anyone can interact with them in real time. Operators from OpenAI, Anthropic, DeepMind, xAI, Cognition, and Wiz joined the round personally.
What it means: When people who build frontier models put their own money into a company that explicitly says the model isn't the hard part, that is a signal about where the next wedge is. In physical AI, the defensible layer may be the boring one — the interface, the deployment, the making-it-usable — not another foundation model. If you're picking a wedge in a hardware-adjacent space, that's the read: own the layer that makes the intelligence usable.
Do this before next week: upgrade to the stable MCP SDKs and put your deprecation migrations on a calendar instead of a fire drill; if an open-weights decision was stuck on political risk, unstick it; and if you're near physical AI, ask whether the usable-interface layer is a better wedge than the model layer. Two of this week's three moves reward planning over speed — which, for once, favors the small team.
