{
  "id": "https://agentreport.io/cross-app-agents/",
  "slug": "cross-app-agents",
  "title": "Agents are crossing app boundaries. The handoff matters.",
  "description": "A source-based briefing on Google Workspace and Claude: where tasks start, where work happens, and what users still need to verify.",
  "url": "https://agentreport.io/cross-app-agents/",
  "markdownUrl": "https://agentreport.io/articles/cross-app-agents.md",
  "jsonUrl": "https://agentreport.io/articles/cross-app-agents.json",
  "format": "briefing",
  "type": "Briefing",
  "published": "2026-10-07",
  "modified": "2026-10-07",
  "author": {
    "name": "Agent Report",
    "kind": "house"
  },
  "status": "Source review and analysis — not a comparative test",
  "disclosure": "This briefing connects announcements from March and September 2026. Neither establishes which product would complete your task more reliably, and we have not compared them hands-on.",
  "language": "en",
  "topics": [
    "interoperability",
    "personal-assistants"
  ],
  "entities": [
    {
      "name": "Google Workspace",
      "type": "SoftwareApplication",
      "url": "https://workspace.google.com/"
    },
    {
      "name": "Claude",
      "type": "SoftwareApplication",
      "url": "https://claude.ai/"
    }
  ],
  "sections": [
    {
      "title": "Google: create the next artifact from the current app",
      "url": "https://agentreport.io/cross-app-agents/#google-create-the-next-artifact-from-the-current-app"
    },
    {
      "title": "Anthropic: assign from a phone, execute on a computer",
      "url": "https://agentreport.io/cross-app-agents/#anthropic-assign-from-a-phone-execute-on-a-computer"
    },
    {
      "title": "Our reading: the product is the whole handoff",
      "url": "https://agentreport.io/cross-app-agents/#our-reading-the-product-is-the-whole-handoff"
    },
    {
      "title": "What to try first",
      "url": "https://agentreport.io/cross-app-agents/#what-to-try-first"
    }
  ],
  "sources": [
    {
      "title": "September 9 Workspace announcement",
      "url": "https://workspace.google.com/blog/product-announcements/less-switching-more-flow-5-new-agentic-capabilities-across-google-workspace-apps",
      "publisher": "workspace.google.com"
    },
    {
      "title": "March 23 announcement",
      "url": "https://claude.com/resources/articles/dispatch-and-computer-use",
      "publisher": "claude.com"
    }
  ],
  "minutes": 3,
  "markdown": "# Agents are crossing app boundaries. The handoff matters.\n\nA source-based briefing on Google Workspace and Claude: where tasks start, where work happens, and what users still need to verify.\n\nCanonical: https://agentreport.io/cross-app-agents/\nAuthor: Agent Report\nPublished: 2026-10-07\nModified: 2026-10-07\nEvidence: Source review and analysis — not a comparative test\nDisclosure: This briefing connects announcements from March and September 2026. Neither establishes which product would complete your task more reliably, and we have not compared them hands-on.\n\nThe starting screen tells you less about an agent than the work it can carry through to another application. Two vendor announcements illustrate that shift. They also expose a practical question: when a task leaves the conversation, how do you know what happened?\n\nThis briefing connects announcements from March and September 2026. It is a reading of those sources, not breaking news or a hands-on comparison.\n\n## Google: create the next artifact from the current app\n\nIn its [September 9 Workspace announcement](https://workspace.google.com/blog/product-announcements/less-switching-more-flow-5-new-agentic-capabilities-across-google-workspace-apps), Google described Gemini handling tasks across Gmail, Drive, Docs, Slides and Chat. Examples include creating a document from an email thread and drafting a presentation from a conversation.\n\nGoogle says the work can draw on selected sources and sources enabled by an administrator. Its email example includes a draft for review and recipient selection before sending. The announcement listed a rollout to several paid Workspace and Google AI plans; it should not be read as proof that every account has every feature today.\n\n## Anthropic: assign from a phone, execute on a computer\n\nAnthropic’s [March 23 announcement](https://claude.com/resources/articles/dispatch-and-computer-use) introduced computer use alongside Dispatch, which lets users assign tasks from a phone. The page describes Claude using connectors when available and screen interaction when needed.\n\nThe same source calls computer use a research preview, notes that it can make mistakes, and says the desktop application must remain awake and running. Those operating conditions matter as much as the demonstration: the phone can be the place a task begins while another machine remains responsible for execution.\n\n## Our reading: the product is the whole handoff\n\nThese are different approaches. One follows the context inside a suite of work applications. The other lets the conversation reach work on a computer. Neither announcement establishes which product would complete your task more reliably.\n\nFor a useful trial, follow the task through four points:\n\n1. **Instruction.** Did the agent preserve the constraints you gave it?\n2. **Access.** Which files, applications and accounts supplied the context?\n3. **Decision.** Where did you review the proposed action before it affected someone else?\n4. **Receipt.** Can you open the actual output and verify the result?\n\nA polished chat response can conceal a missing attachment or an unfinished action. A plain response with a correct, inspectable artifact may be more useful. That is why we will evaluate the destination as well as the conversation.\n\n## What to try first\n\nChoose an existing document or a small set of non-sensitive files. Ask for one concrete output with a clear destination. Keep the first run bounded enough that you can compare the result with the source yourself.\n\nRecord what you had to correct and which steps remained manual. That gives you a practical basis for deciding whether a broader connection or a recurring task is worthwhile.\n\nSee [our testing method](https://agentreport.io/how-we-test/) for how those observations become a report. Browse [Agentlist](https://www.agentlist.io/) when you are ready to compare the wider field.\n",
  "text": "Agents are crossing app boundaries. The handoff matters. A source-based briefing on Google Workspace and Claude: where tasks start, where work happens, and what users still need to verify. Canonical: https://agentreport.io/cross-app-agents/ Author: Agent Report Published: 2026-10-07 Modified: 2026-10-07 Evidence: Source review and analysis — not a comparative test Disclosure: This briefing connects announcements from March and September 2026. Neither establishes which product would complete your task more reliably, and we have not compared them hands-on. The starting screen tells you less about an agent than the work it can carry through to another application. Two vendor announcements illustrate that shift. They also expose a practical question: when a task leaves the conversation, how do you know what happened? This briefing connects announcements from March and September 2026. It is a reading of those sources, not breaking news or a hands-on comparison. Google: create the next artifact from the current app In its September 9 Workspace announcement, Google described Gemini handling tasks across Gmail, Drive, Docs, Slides and Chat. Examples include creating a document from an email thread and drafting a presentation from a conversation. Google says the work can draw on selected sources and sources enabled by an administrator. Its email example includes a draft for review and recipient selection before sending. The announcement listed a rollout to several paid Workspace and Google AI plans; it should not be read as proof that every account has every feature today. Anthropic: assign from a phone, execute on a computer Anthropic’s March 23 announcement introduced computer use alongside Dispatch, which lets users assign tasks from a phone. The page describes Claude using connectors when available and screen interaction when needed. The same source calls computer use a research preview, notes that it can make mistakes, and says the desktop application must remain awake and running. Those operating conditions matter as much as the demonstration: the phone can be the place a task begins while another machine remains responsible for execution. Our reading: the product is the whole handoff These are different approaches. One follows the context inside a suite of work applications. The other lets the conversation reach work on a computer. Neither announcement establishes which product would complete your task more reliably. For a useful trial, follow the task through four points: 1. Instruction. Did the agent preserve the constraints you gave it? 2. Access. Which files, applications and accounts supplied the context? 3. Decision. Where did you review the proposed action before it affected someone else? 4. Receipt. Can you open the actual output and verify the result? A polished chat response can conceal a missing attachment or an unfinished action. A plain response with a correct, inspectable artifact may be more useful. That is why we will evaluate the destination as well as the conversation. What to try first Choose an existing document or a small set of non-sensitive files. Ask for one concrete output with a clear destination. Keep the first run bounded enough that you can compare the result with the source yourself. Record what you had to correct and which steps remained manual. That gives you a practical basis for deciding whether a broader connection or a recurring task is worthwhile. See our testing method for how those observations become a report. Browse Agentlist when you are ready to compare the wider field."
}
