Our reading, which is interpretation, is that the single-loop agent guidance and Claude Duet show why Anthropic wants to run this layer itself.
Summary
Anthropic made Claude Models on Google Cloud (Anthropic managed) generally available. Google Cloud customers get the full Claude Platform with Google IAM credentials and one Google invoice, and the usage counts toward their existing cloud commitments. Anthropic operates the service and is the data processor. In the same cycle, Anthropic published guidance on folding subagents into one agent loop and opened a waitlist for Claude Duet, a background agent that can draft emails, chase invoices and pay for purchases the user approves.
What happened
Google Cloud customers can now get Claude in two ways. The existing Google-managed path stays in place for customers who need Google as data processor, regional data residency, or integrations such as Model Armor and Agent Engine. The new Anthropic-managed path is pitched as the full Claude Platform, with features and betas arriving the same day they ship natively. Anthropic lists Claude Managed Agents, code execution, computer use, agent skills, the Files API, citations, web search and web fetch, an MCP connector, prompt caching and batch processing. It runs on US and Global endpoints. Fast Mode for Claude Opus 5.5 is generally available on this path at $8 per million input tokens and $40 per million output tokens, and Anthropic says it produces output tokens up to 2.5x faster. The available models are Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5.
The agent-architecture post argues that many setups with routers, planners and domain subagents can be collapsed into one loop. That means one model and one context window, with skills and tools loaded only when needed. The harness keeps turn and cost budgets, write gates and optional verifier checks. Anthropic cites one customer comparison where the single loop ran about twice as fast. It also says progressive disclosure cut Claude Code's system prompt by more than 80% for advanced models. The post keeps subagents for heavy context pressure, parallel work with disposable context, hard tool selection and conflicting instructions.
Duet is a research preview for US Pro and Max subscribers. It runs on Opus 5.5 in a cloud VM with its own browser and uses the user's memory, skills and connectors. Anthropic cites more than 6,000 Marketplace connectors and plugins. Duet can pay with Stripe Link one-time cards for purchases the user approves. It asks before sending messages, making payments, signing documents or deleting data. Securities trading and transfers between accounts are unavailable even with approval.
What the archive adds
The dossier fills in how the Google Cloud arrangement works. According to Anthropic's documentation (d2), Anthropic meters the usage. Google Cloud Marketplace reports it in Claude Consumption Units, it appears on the Google Cloud bill, and it counts toward the customer's spend commitment. Google's own documentation (d3) describes the Google-managed path as serverless Claude APIs inside Gemini Enterprise Agent Platform, with pay-as-you-go or provisioned-throughput pricing and Google Cloud's FedRAMP High boundary. The Anthropic-managed documentation excerpted here does not cover FedRAMP, residency or Model Armor. That silence is not evidence that the controls are missing; the excerpts simply do not answer the question. Earlier archive coverage of Duet describes it as built "for professionals and small business owners". That places it as a prosumer product, not an enterprise one.
Trend signal
The pattern is a model vendor pulling the agent runtime back from cloud resellers while keeping the cloud's procurement benefits. Most of the features Anthropic lists for the new path belong to the runtime, not the model: Managed Agents, skills, the MCP connector, code execution and computer use. They depend on Anthropic's harness and change often. A same-day-betas promise only makes sense if Anthropic runs the service itself. That is the structural difference between this path and the Google-managed one. Our interpretation is that the other two stories point the same way. The single-loop post says the gains come from managing the loop well, through progressive disclosure, budgets, write gates and verifiers, and those are things Anthropic's platform provides. Duet applies the same set of pieces to consumers: memory, skills, connectors, a cloud computer and gated payments. Anthropic does not present these three as one strategy. That framing is ours.
Why now
1) Commitment drawdown. According to d2, Anthropic meters usage, Google Cloud Marketplace reports it in Claude Consumption Units, and it counts toward Google Cloud spend commitments. That gives Anthropic access to budgets customers have already committed. 2) Release control. The primary source's main promise is same-day features and betas, which only the operator of the service can guarantee. 3) Agent features belong to the runtime. Managed Agents, skills and the MCP connector depend on Anthropic's harness, and the single-loop guidance puts that harness at the center of agent design. 4) A premium tier to sell. Fast Mode for Opus 5.5, at $8/$40 per million tokens, launches on the Anthropic-run path.
Connections to other events
Here is how the motive looks in practice. The evidence shows Google keeping the identity layer (IAM), the invoice and the commitment drawdown, measured in Claude Consumption Units (d2). Anthropic keeps the runtime, the data-processor role and the release schedule. Our read is that the split is deliberate. Anthropic gets Google Cloud's billing and procurement channel without giving up control of the product surface where its agent features ship.
The same-day promise matters most for the features the single-loop guidance tells builders to depend on. These are skills loaded progressively, managed agents and write gates at the harness level. A team that designs around them is designing around Anthropic's runtime. On Google Cloud, the path described as receiving those features on release is the one Anthropic runs. Fast Mode is the pricing side of the same story. A premium SKU that trades higher price for speed on the flagship model ships on the Anthropic-run path at $8/$40 per million tokens. The primary source does not say whether Fast Mode is also offered through Google-managed Claude.
Duet is our speculative extension. Background execution, approval gates on payments and signatures, and a large connector catalogue are capabilities that would most plausibly reach enterprise customers first through surfaces Anthropic operates. Nothing in the evidence commits Anthropic to that.
The split also leaves something on Google's side. FedRAMP High (d3), regional residency, Model Armor and Agent Engine are listed for the Google-managed path. For regulated buyers, the choice is currently between those listed controls and the full agent feature set. This is not because the Anthropic-managed path is known to lack those controls. It is because the documentation excerpted here does not address them.
Why it matters
For builders on Google Cloud, the choice of path is now a question of which features you need and which controls you need. If your roadmap depends on managed agents, skills, the MCP connector or new betas, the Anthropic-managed path is the one described as getting them at launch. If you need the controls Google lists, such as FedRAMP High, residency or Model Armor, the Google-managed path is where they are documented today. Until Anthropic documents equivalents, expect the full agent feature set and those controls to sit on different paths. For buyers, d2 shows that committed Google Cloud spend can now fund the full Anthropic platform. For architects, combining the single-loop guidance with same-day betas means building deeper into Anthropic's harness. You get simpler and faster designs but lose portability. Keep evals and tool schemas vendor-neutral so that the loop is the only part you would need to rebuild.
What happens next
- Anthropic-managed Claude on Google Cloud adds at least one named regional endpoint beyond US and Global, such as EU. (6 months, 50%). Watch for: Anthropic or Google Cloud documentation lists a new regional endpoint for the Anthropic-managed offering.
- At least one new Claude Platform beta is documented as supported on the Anthropic-managed Google Cloud path but missing from Google's partner-model Claude documentation at launch. (6 months, 70%). Watch for: Release notes name Anthropic-managed Google Cloud as supported, while Google's Claude partner-model page omits the feature.
- Claude Duet opens to Team or Enterprise plans, or Claude Platform exposes background agent execution with approval gates. (12 months, 45%). Watch for: The Help Center removes the 'Team and Enterprise not supported' line, or Managed Agents documentation describes scheduled or background runs with approval gates.
Limits and uncertainty
The excerpts here contain no feature-by-feature matrix for the two Google Cloud paths. The claim that agent-runtime features are exclusive to the Anthropic-managed path is inferred from the "full platform" and same-day framing, not from an explicit list. The announcement does not say which discounts or provisioned-throughput terms carry over, what regions exist beyond US and Global, or how latency compares with the native API. The motive is inferred from how the offer is structured, not from anything Anthropic said, and Google may have pushed for this arrangement itself. The 2x speedup and the 80% prompt reduction come from Anthropic's own selected examples. Duet is a US-only preview with no announced pricing after the preview, so linking it to the enterprise strategy is interpretation. It is also unclear how audits treat two Claude deployments with different data processors inside one cloud account.
Skill to build now
Build an end-to-end task eval that does not depend on your architecture. This week, write 10 golden examples from one real workflow. Give each an input, an expected end state and a pass/fail check. Run your current subagent design against them, then a single-loop version that loads skills only when needed. Record completion rate, wall-clock time and token cost for each. You can reuse the same set to compare the Anthropic-managed and Google-managed paths, or Fast Mode against standard Opus, without rewriting anything.
Reader takeaway
If you run Claude on Google Cloud, list two things for each workload: the platform features your agents depend on (managed agents, skills, MCP connector, code execution) and the controls you are required to have (residency, FedRAMP High, Model Armor). Move one feature-heavy, non-regulated workload to the Anthropic-managed path and confirm that its usage draws down your commitment. Keep regulated workloads where their controls are documented. Ask Anthropic in writing about compliance coverage before moving anything else.
Open question
Now that Anthropic's own runtime is sold through Google's billing, will Google-managed Claude gain the agent-runtime features, or will it settle into being the compliance path while Anthropic holds the product relationship with Google's customers?