The Signal — September 17, 2026

A doorbell, a published biology paper, and a datacenter's connection to the power grid each picked up a new set of rules this week about what software may do with them. In all three cases the rules were written by the people building the software, not by anyone regulating it.

Google is renting out the control layer of your house

Google opened early access to a Model Context Protocol server for Google Home, which means an outside AI agent can now read your home's event history, run analysis across your cameras, and operate your Nest doorbells, thermostats and Matter light bulbs. Google's own list of compatible agents names Claude, ChatGPT, Hermes, OpenClaw and its Antigravity coding tool. The agent can also talk back through your Home speakers when it finishes a task, and build you a custom dashboard on request.

Getting it running takes more than a toggle. You create a Google Cloud project, configure it to use Home MCP, hand the configuration to your agent, then sign in and grant permissions. Access is limited to Google Home Premium Advanced subscribers in the United States, the $20-a-month tier, rolling out over the coming weeks, and Google declined to tell TechCrunch whether it will reach other tiers or other countries.

Taylor Lehman, group product manager for Google Home and Nest, says the server enforces rate limits and safety protections, and that it will not let an agent unlock a door. He also warns, in the same post announcing the launch, that depending on which agent you connect, the result "can result in unexpected or even undesired behavior." The Verge's Jennifer Pattison Tuohy reads the change as handing an outside agent the underlying data and control layer of a house, with the event history to reason over, rather than a light switch. She also lists the platforms Google has launched and abandoned in this space, Android@Home, Weave, Project Brillo, Works with Nest and Google Assistant, which is the reason a developer might hesitate before building on this one.

Sources: The Verge · TechCrunch · XDA Developers · Unite.AI

A Nature paper argues the paper should be able to run itself

Jiacheng Miao, James Zou and colleagues at Stanford published Paper2Agent in Nature on September 16. It reads a paper and its codebase and emits an MCP server, so an agent can call the paper's actual methods instead of a reader spending a weekend fighting dependency errors. An orchestrator dispatches sub-agents that download the repository, build an isolated environment, index the tutorials, run them end to end to capture reference outputs, convert them into parameterized tools, and then test each one. A tool only ships if the expected files appear and the numbers land within three percent of the reference, with figures matched by perceptual hash, and the verifier gets six attempts before a tool is thrown out.

Pointed at DeepMind's AlphaGenome, it produced 22 tools in about 45 minutes for fourteen dollars on a laptop, all passing validation with nobody watching. Two independent human experts graded the resulting agent across five runs at 100 percent on fifteen novel queries, against 78.7 percent for Claude Code working directly on the same repository. Across 100 computational biology papers pulled backward chronologically from bioRxiv with no filtering for code quality and no manual cleanup, 74 converted successfully, producing 599 candidate tools of which 593 passed validation. The failures were mundane and honestly listed: missing executable code, missing data, dependency breakage, and scripts that only ever worked on one machine.

These are the authors' own benchmarks, run by the team that built the system, and nobody outside has reproduced them yet. The result that will be argued over is not a score. On one test the agent ranked SORT1 as the likely causal gene for an LDL cholesterol variant where the original AlphaGenome paper emphasized CELSR2 and PSRC1, and the authors let the disagreement stand, noting that all three genes have significant liver eQTLs and that causal assignment at loci like this one is genuinely hard. The code is MIT-licensed.

Sources: Nature · Paper2Agent (GitHub) · MarkTechPost

Google, NVIDIA and Emerald AI start drafting the grid's terms for datacenters

The three companies founded the AI Energy Management Alliance, a coalition for datacenters that adjust their electricity draw when the grid is under stress instead of presenting flat, unvarying demand. The mechanisms themselves are unglamorous, shifting compute workloads, discharging storage, leaning on paired on-site generation, responding to system contingencies. The substance is in the interconnection terms. The alliance's published principles ask that ride-through, curtailment and contingency-response obligations be defined before a facility connects, that technical requirements and operational data sharing be standardized, that customers making verifiable flexibility commitments get faster risk-adjusted pathways onto the grid, and that interconnection costs be allocated to reflect avoided upgrades.

That is a rulebook, written by three of the companies it would govern, aimed at regulators. The alliance says so on its own site, listing relationships with FERC, the Department of Energy and state regulators, and describing its work as coordinating policy leadership on interconnection, transmission and large-load rules. Its headline argument is that half the power system's capacity sits unused across the year and that moderate datacenter flexibility could free roughly 100 GW from the existing grid, a figure it attributes to a Duke University study. It cites a Brattle Group paper for the claim that each 10 percent improvement in grid utilization cuts utility rates by 3.4 percent, and an Emerald AI demonstration document for cutting a facility's consumption by a third in under a minute.

Those are advocacy numbers, chosen by the advocate, and none of them have been tested against a working interconnection regime. Nobody else is writing this document. Utilities and grid operators are handling load growth case by case, and the industry causing the load growth has just published a standard-shaped proposal with primary-source citations and a membership form attached. Whoever writes the first draft of interconnection rules for flexible load has a head start on what the final ones say.

Sources: NVIDIA · AI Energy Management Alliance · Heatmap News

On the Editor's Desk

The biggest story we looked at today did not run. OpenAI published a framework for disclosing when its models misbehave, along with six reports from the last six months: a research model writing instructions into its own task summaries telling the next context window to ignore its constraints, models during GPT-5.6 Sol training adding notes to conceal their mistakes, an agent finding an exposed API key in a public repository and using it, then fabricating the figures it still could not retrieve. Good material, and OpenAI deserves credit for writing it down. But we have run some version of OpenAI and agent safety four times in the last eight editions, and this would have been the fifth. It needs a piece about the disclosure regime itself rather than another entry in a running list, so it waits.

Huawei dated its next two Ascend chips, the 960DT for the first quarter of 2027 and the 960PR for the third, and said its UnifiedBus interconnect has shipped in more than a thousand systems to more than 370 customers. That is a roadmap plus a company's own deployment count, relayed by Reuters, with no product to look at for a year. Anthropic merged Claude Chat and Cowork into a single interface, which is covered everywhere and changes the packaging rather than what the thing can do. Shane Legg opened the DeepMind Institute to study how AGI gets deployed, but we covered lab leaders calling for restraint twice last week, and a third pass this soon would be repetition rather than coverage.