Anthropic has introduced Claude Duet, an AI agent that does not wait for input but continues working independently – even when the user is not at the computer at all.The technical basis is the Claude Opus 5.5 model.Via around 6,000 connectors, i.e. pre-built interfaces to external services, the application accesses systems such as Gmail, Slack or Stripe.Before sensitive steps – payments are cited as an example – the agent requests explicit approval.

Anthropic is not alone in this. Similar products come from xAI, Meta and OpenAI.The market is therefore visibly moving away from the chat window, in which every action is triggered individually, and towards agents that run continuously and decide for themselves when to become active.

From assistant to continuous runner

The decisive difference from previous assistance systems lies less in the language model than in the operating mode.A classic chatbot is reactive: a question goes in, an answer comes out, and after that nothing more happens.A proactive agent, by contrast, remains active in the background, monitors connected systems and triggers actions itself.

This also shifts the character of the integration.As long as a model only produces text, the damage caused by an error remains manageable.As soon as the same model sends emails, posts messages in a team channel or triggers transactions in a payment system, every misjudgement has an immediate external effect.The connectors are therefore the actually critical part of the architecture, not model quality alone.

Approvals as an architectural decision

That Claude Duet asks for confirmation before sensitive actions is conceptually obvious – in practice, the difficulty lies in drawing the boundary.What counts as sensitive?A payment is a clear case.An email to an existing customer, a status change in a ticket system or the creation of a record in the CRM are not automatically so – but in individual cases they can be just as consequential.

Companies that want to deploy such agents must draw this boundary themselves.It makes sense to sort actions by direction of effect:

  • Reading: Retrieve, summarise and evaluate data.Here the risk is primarily a data protection risk, not a process risk.
  • Writing internally: Notes, drafts, tickets, internal messages.Errors remain correctable and visible.
  • Acting externally: Customer communication, orders, payments, contract documents.Here human approval belongs in the standard case, not the exception.

If this classification is not made in advance, the provider takes it over with its default settings.That may be sufficient for individual users, but it is rarely enough for a company with compliance requirements.

Guardrails are more than a confirmation dialog

A confirmation window before payment is only one of several protective layers.Anyone deploying agents in production additionally needs traceable logging: which action was executed when, by which trigger, on what data basis, with which approval?Without this log, an error cannot be reconstructed after the fact – and that is precisely what is required during audits.

In addition, there is the allocation of rights.An agent that works with an account that covers all mailboxes and payment transactions concentrates an enormous amount of agency at a single point.The principle of least privilege – each account only to the extent necessary for the specific task – applies here just as it does to human user accounts, but is gladly skipped with quickly set-up connectors.

A third point concerns persistence.An agent that continues working even when offline potentially runs at night, at weekends and during holiday periods.Time limits, caps on actions per period and a reliable emergency stop are therefore part of the basic equipment.

Where the benefit realistically lies

The appeal of these systems lies in processes that currently cost a lot of context switching: gathering information from multiple tools, pre-sorting transactions, preparing drafts, reporting deviations.These are precisely the activities that in mid-sized organisations often depend on individual people and are left undone in their absence.

Less realistic is the expectation that an agent will take over a business process completely and without supervision.Processes contain exceptions, tacit rules and discretionary scope that are documented nowhere.Anyone who does not make them explicit beforehand merely shifts them into a system that resolves them according to its own logic – and the result then sometimes turns out plausibly worded but factually wrong.

Dependence and interchangeability

The figure of around 6,000 connectors shows what the competition between providers is actually about: the breadth of connectivity.The more systems an agent can reach, the more work it can take over – and the more deeply it grows into one's own infrastructure.

For planning purposes, it is therefore worth looking at exit costs.If workflows, prompts and approval rules exist exclusively in a provider's interface, a later switch is costly.

It makes more sense to document the technical logic – i.e. the description of the process, the approval stages and the review criteria – independently of the specific product and to treat the AI component as a replaceable module.

Classification for web and software projects

For project practice, the focus shifts: the actual work no longer lies in connecting a model, but in properly designing permissions, approval stages and logging. Interfaces to existing systems should be built from the outset in such a way that they can also be used by an automated counterpart – with tightly scoped rights and full traceability. Anyone who starts now with clearly delimited, low-risk use cases gains the necessary experience before agents reach into processes with a real external impact.

Sources