Should A Cloud Agent Run Your Business, Or Should You Keep It On-Prem?

Last Updated: 23 October 2026

Key takeaways

• Claude Duet is Anthropic's proactive agent, running on Opus 5.5 inside a cloud computer, and as of 23 October 2026 it is a waitlist research preview for Pro and Max subscribers in the United States only.
• A cloud agent is the right tool for work that is public, reversible and cheap to check: morning briefs, drafts, research, scheduling and summarising.
• An on-prem agent is the right tool for work that touches customer records, payroll, pricing, contracts or credentials, where the data must not leave your control.
• Flowtivity runs its own agent on its own hardware: DeepSeek V4 Flash on two DGX Sparks, measured at 35 tokens per second on a single stream and 62 tokens per second at four concurrent requests.
• The decision is not cloud or local. It is a per-task decision you make once, write down, and review monthly.
• Under Australian Privacy Principle 8, sending personal information to an overseas recipient is a disclosure you have to be able to justify.

Short answer: let a cloud agent run the parts of your business that are public, reversible and cheap to check, and keep the parts that touch customer records, payroll, pricing, contracts and credentials on hardware you control. Anthropic opened Claude Duet as a US-only research preview on 21 October 2026, and the whole point of the product is that it keeps working after you close the app. That changes the risk conversation for Australian small businesses. The question is no longer whether an agent can run your business. It is which tasks you are willing to hand to a computer you do not own, and which you insist on running yourself.

What Claude Duet actually is

Claude Duet is a proactive agent that runs on Opus 5.5 and works inside its own cloud computer. According to Anthropic's product page, it briefs you every morning, takes tasks you hand it by message, keeps working while you are away, and messages you when a task is done or needs a decision. It works with more than 6,000 connectors and plugins, and it carries your existing Claude memory, skills and apps into the first conversation. It is a standalone app on iOS and Android, plus claude.ai on a computer. The availability caveat matters: Duet is a research preview on a waitlist for Pro and Max subscribers in the US, and Anthropic says Team and Enterprise plans are not included yet.

One tester described the environment in the launch thread on Hacker News, which drew 550 points and 516 comments on its first day: each session runs in its own cloud Linux container with a shell, code tools and a browser, so the agent can run scripts and work through websites rather than only calling APIs. That is the same architecture pattern as an on-prem agent. The difference is who owns the container.

What a cloud agent will not do without you

Anthropic has built approval gates into Duet rather than leaving them to the user. According to the product page, Claude Duet asks before sensitive operations such as sending messages or paying, and approvals, logins and safety checks run inside your cloud computer so sensitive details stay private. You can also see what it has done and is doing, and steer it. The same page is refreshingly blunt about the gaps: it may misread a thread, miscalibrate what is urgent, and miss something that arrives between check-ins. That is a fair description of any agent, cloud or local.

The trust problem shows up in the user reaction, not the marketing. In the Hacker News launch thread, one commenter wrote that agents "cannot intuit what actions they can take on my behalf that I will or will not approve of", which is why he will not use one. Another described buying a 128GB Spark to run an agent in house for a family of four, and noted that high tokens per second is not the bottleneck for tasks like booking seats together at a cinema. Both comments point at the same thing: the hard part of an agent is not intelligence, it is the boundary around what it may touch.

The four limits every cloud agent hits in a small business

Every hosted agent, not just Duet, runs into four limits that matter more to a 20-person business than to a solo professional. The first is data residency: the preview is US-only and processing happens in Anthropic's cloud, so customer records, supplier pricing and payroll data leave Australia. The second is lock-in: memory, skills and connectors live in one account tied to one model family, and moving them later is a migration project. The third is cost predictability: Duet shares your plan's usage limits with the rest of Claude, and its background checks count toward those limits, so a heavy day of proactive work quietly consumes the same budget as your chat. The fourth is credential spread: 6,000 connectors means 6,000 tokens and keys, and each one is an opening into a system you own.

The regulatory piece is not theoretical. Under the Privacy Act 1988, Australian Privacy Principle 8 requires an organisation to take reasonable steps to ensure that an overseas recipient does not breach the Australian Privacy Principles before it discloses personal information to them. Using a US-hosted agent on customer records is a cross-border disclosure, and the OAIC expects you to be able to explain the basis for it. For a business holding health information, financial details or children's data, the simplest defensible answer is often that the data never leaves the building.

What an on-prem agent looks like in practice

An on-prem agent is not a different product category. It is the same loop, the same tools and the same approval gates, running on hardware you own. At Flowtivity we run one every day: an agent framework on our own server, a Telegram interface for the humans, shell, file and browser tools, scheduled jobs, and an allow-list that stops it touching anything we have not approved. The model is DeepSeek V4 Flash on two DGX Sparks with 128GB of unified memory each. In our own testing it sustained 35 tokens per second on a single stream and 62 tokens per second at four concurrent requests, which is slow by API standards and fast enough for agent work, because agent work is dominated by tool calls and waiting, not by token generation.

The practical difference is where the data sits. Customer lists, quote history, payroll files and contract drafts stay on our network. The agent can read them because it runs next to them. When a task needs frontier reasoning, we can send a summary or a redacted excerpt to a hosted model and keep the underlying records local. That hybrid is not a compromise. It is the design that survives an audit.

Cloud agent vs on-prem agent, side by side

The comparison is not about capability, because the same model can sit behind both. It is about custody. A cloud agent gives you a faster start and someone else's maintenance burden. An on-prem agent gives you control of the data and the keys, and hands you the operational work. Most businesses end up with both, and the useful question is which task belongs where.

Which tasks go where: a three-bucket rule

Sort every task into one of three buckets before you give an agent access to anything. Bucket one is cloud-safe: public information, first drafts, research, calendar handling and summaries of documents that are already shareable. Bucket two is hybrid: sensitive data summarised locally, with only the summary sent to a hosted model, or a local model drafting while a person approves. Bucket three is never-cloud: customer records, payroll, pre-signature contracts, credentials, health information and anything under an NDA. The rule is not a judgement about which model is smarter. It is a judgement about which data you are willing to explain to a regulator, a client or a court.

What it actually costs

The cloud side is cheap and predictable at the low end. According to Anthropic's pricing page, Claude Pro is US$20 per month, or US$17 per month billed annually, and Max runs from US$100 to US$200 per month depending on usage multiplier. API access to Opus 5.5 is US$4 per million input tokens and US$20 per million output tokens. On subscription maths alone, a hosted agent wins for a single user, and it wins clearly. The on-prem side carries a real bill: NVIDIA Marketplace listed the 128GB DGX Spark at about US$6,950 in October 2026, up from its US$3,999 launch price, and two units draw roughly 240 watts each, which is about A$100 per month in power at Queensland rates.

So do not buy hardware to save subscription money. Buy it when one of four things is true: several staff need the same agent, you are running API workloads at a scale where per-token billing hurts, the data cannot legally or commercially leave your control, or you want the agent to do work a hosted plan will not permit. In our case it was the last two. The hardware also does non-agent work, which is how the maths usually clears.

How to start without betting the business

Start with the boundary, not the model. Write down what the agent may never touch, then give it a narrow first job and a person who approves every outbound action. The sequence below is the one we use with clients, and it works whether the agent ends up in Anthropic's cloud or on a box in your office. The point of the first month is not autonomy, it is evidence: a log of what the agent did, what it got wrong and what it should never have been allowed to attempt.

1. Write the never list. Name the systems the agent may not touch: payroll, bank accounts, customer databases, password managers, production infrastructure. Put it in the agent's instructions and in the access policy.
2. Start with drafts only. No sends, no payments, no deletions. The agent prepares, a person approves. This is exactly the pattern Anthropic ships with Duet, and it is the right default on any hardware.
3. Split the work by data class. Public and shareable tasks go to the cloud agent. Anything with personal information stays on local hardware or is summarised before it leaves.
4. Log every action and review weekly. Tool calls, files touched, messages drafted. A weekly sample review catches drift long before a client does.
5. Re-check the split monthly. Tasks migrate from cloud to local as volume grows, and from local to cloud as models improve. The split is a living decision, not a one-off.

Frequently asked questions

Is Claude Duet available in Australia?

No. According to Anthropic's product page, Claude Duet is a research preview on a waitlist for Pro and Max subscribers in the United States, and Team and Enterprise plans are not included yet. Australian users can use Claude, Claude Code and the API, but not Duet as of 23 October 2026.

What does Claude Duet do that ordinary Claude does not?

Claude waits for you to open a chat. Duet comes to you: it sends a morning brief, takes tasks by message, keeps working while you are away, and messages you when something is finished or needs your approval. Anthropic describes it as a proactive agent rather than an assistant you drive.

Will a cloud agent send emails or make payments on its own?

Anthropic says Duet asks before sensitive operations like sending messages or paying, and that approvals, logins and safety checks run inside your cloud computer. Testers still report misjudged urgency and missed items between check-ins, so the approval gate is a safety net rather than a guarantee.

Can I run a proactive agent on my own hardware instead?

Yes. An open-source agent framework plus an open-weight model on a machine with enough unified memory gives you the same loop: scheduled jobs, tool calls, approvals and a message interface. Flowtivity runs one on two DGX Sparks with DeepSeek V4 Flash, measured at 35 tokens per second on a single stream and 62 at four concurrent requests.

Is sending customer data to a US cloud agent a privacy problem in Australia?

It can be. Australian Privacy Principle 8 requires reasonable steps to ensure an overseas recipient does not breach the Australian Privacy Principles before personal information is disclosed to them. If the agent processes customer records in the US, you need a documented basis for that disclosure, and for some data classes the cleanest answer is to keep it local.

The short version

Claude Duet is a good product and a genuine shift: the agent now comes to you, works while you are away, and asks before it spends money or sends a message. It is also a US-only preview that puts your business context in someone else's cloud, shares your plan limits with your chat usage, and asks you to trust 6,000 connectors. That is a reasonable trade for drafts, briefs and research. It is a poor trade for payroll, customer records and contracts. The durable answer for an Australian small business is a split: hosted agents for public, reversible work, an on-prem agent for the data you cannot send offshore, and a written rule that says which is which. You do not have to choose a side. You do have to choose per task, and write it down before the agent does something you cannot undo.