AI briefing: agents move into workspaces, homes and ad systems

Key developments from the 24 hours to 17 September 2026 UTC centre on where AI agents can act: inside productivity software, smart homes, advertising systems and power-constrained data centres.

1. Anthropic merges Claude Cowork into chat and adds Docs and Slides

Anthropic is combining its Cowork agentic workspace with the main Claude chat interface. The change brings longer-running tasks, connectors and skills into ordinary conversations, while new Claude Docs and Claude Slides tools let paid users create and edit documents and presentations. The rollout begins for Pro and Max users over the coming weeks; the new creation tools are in beta.

Why it matters: This removes a product boundary between conversational assistance and delegated work. It also makes Claude a more direct competitor to office-suite AI products by keeping research, drafting, presentation creation and agent execution in one workflow.

Sources

• Claude Cowork and chat are now one Claude — Anthropic — 2026-09-16
• Anthropic merges its chat and agentic products into one AI assistant in push to build a superapp — Fortune — 2026-09-16

2. OpenAI tests business-sponsored agents inside ChatGPT advertising

OpenAI has begun a limited US test of Sponsored Agents: clearly labelled business agents that users can open after clicking an advert in ChatGPT. The company also announced prompt-based ad creation in ChatGPT Work, expanded AI creative tools in Ads Manager, and integrations with HubSpot and Shopify.

Why it matters: The test moves advertising from static targeting towards interactive sales and support conversations. Its separation from ChatGPT’s independent answers will be central to whether users, advertisers and regulators view the format as trustworthy.

Sources

• Reimagining advertising with AI — OpenAI — 2026-09-16
• Sponsored Agents in ChatGPT Ads — OpenAI Help Center

3. OpenAI formalises public reporting of model-misalignment findings

OpenAI published a framework for tracking, investigating and disclosing model-misalignment cases, alongside six reports of unexpected or concerning behaviour observed over the previous six months. It says reports may be published before a full explanation or mitigation is available, and describes the framework as a starting point rather than an industry standard.

Why it matters: The change creates a more regular channel for evidence about agent behaviour and safeguards, rather than leaving such findings mainly to model system cards or occasional retrospective reports. The usefulness will depend on the consistency, detail and independence of future disclosures.

Sources

• Our framework for reporting model misalignment — OpenAI — 2026-09-16
• OpenAI flags concerning new AI behavior and vows to track it more closely — Associated Press — 2026-09-17

4. Google opens early access to Home MCP for third-party agents

Google has opened early access to a Model Context Protocol server for Google Home. The integration is intended to let compatible AI agents interact with smart-home devices and selected home context through Google Home, extending agent access beyond Google’s own assistant experiences.

Why it matters: Giving external agents a bridge into cameras, speakers and connected devices raises the practical value of MCP, but also elevates the stakes for permissions, identity controls and auditability when an AI system can affect the physical environment.

Sources

• Introducing Home MCP: enabling your agent to interact with your home — Google Home & Nest Community — 2026-09-16
• Coverage of Google Home MCP early access — Techmeme — 2026-09-17

5. MLPerf adds agentic inference tests as Nvidia posts Vera Rubin preview results

MLCommons released MLPerf Inference v6.1, adding benchmarks for agentic inference and vision-language workloads. Nvidia’s first Vera Rubin NVL72 preview submission reported up to 3.7 times the throughput of its GB300 NVL72 on Qwen3-VL; MLCommons said the broader release included peer-reviewed results for new platforms and record participation.

Why it matters: Inference cost and latency are becoming defining constraints for reasoning and tool-using systems. Standardised agent-oriented tests are an early step towards comparing the infrastructure that will run those workloads, although vendor-specific headline comparisons still require careful reading of benchmark settings.

Sources

• MLCommons Sets Participation Record with New MLPerf Inference v6.1 Benchmark Results — MLCommons — 2026-09-16
• NVIDIA Vera Rubin NVL72 Delivers Leading Performance in MLPerf Inference v6.1 Debut — NVIDIA — 2026-09-16

6. Google, Nvidia and Emerald AI form flexible-data-centre power alliance

Google, Nvidia and Emerald AI launched the AI Energy Management Alliance, a coalition focused on data centres that can alter electricity demand in response to grid conditions. The founders say the group will develop technical and operational approaches, work with utilities on interconnection, and advocate for policies recognising grid-responsive demand; partners include Anthropic, utilities and power producers.

Why it matters: Power availability is increasingly the limiting factor for AI expansion. If flexible workloads can earn faster grid connections while reducing pressure during peak demand, the approach could affect where and how quickly new AI capacity is built.

Sources

• Emerald AI, Google and NVIDIA Launch Alliance to Advance Flexible AI Data Centers — NVIDIA — 2026-09-16
• Tech giants launch flexible-power coalition for data centers — Axios — 2026-09-16

What to watch

Watch for details on access controls and supported agents in Google Home MCP, the first advertiser outcomes from OpenAI’s Sponsored Agents trial, and whether OpenAI’s new misalignment reporting produces timely, technically substantive disclosures. On infrastructure, follow whether utilities give flexible AI data centres preferential interconnection treatment and how MLPerf’s new agentic benchmark is adopted by competing hardware vendors.