Two announcements from August show the same movement from different directions. Google has launched Gemini Spark, an offering that brings AI agents to a broad audience; it was initially introduced in the US, with release in Germany still announced for this year. Anthropic, in turn, has given Claude an update that enables further AI actions in Google services – including Gmail, where the model can take on tasks relating to emails.

Both are individual product steps. Together, however, they mark a shift that is more relevant to the organisation of work than any new model generation: away from the chat window, in which someone asks a question and receives an answer, towards software that acts on its own within existing applications.

What distinguishes an agent from a chatbot

A chatbot delivers text. An agent – that is, an AI system that independently executes several steps to achieve a specified goal – accesses tools and data and produces results outside the conversation window. The difference sounds technical, but it is above all organisational: the output of a chatbot lands on the desk of a person who checks it. The output of an agent may land directly in the outbox.

This is precisely where the break lies. As long as an AI makes suggestions, the human is the final control point – automatically and without anyone ever having had to define it. As soon as an AI acts, this control point must be deliberately established, otherwise it no longer exists.

Is prompt engineering becoming obsolete?

The question has been in the air since agents started emerging from the niche. It is justified, but probably wrongly framed. What actually loses importance is polishing formulations for a single run: the right role description, the clever subordinate clause, the example in the right format. This manual work is migrating into the products themselves, which increasingly bring context and structure with them.

What does not disappear is the underlying competence: describing a task so precisely that a system can process it correctly without asking back. With agents, this requirement tends to grow, because the result is no longer just a text but an action with consequences. Prompt engineering becomes task definition – with a description of the goal, abort conditions, permitted data sources and the specification of what the agent must under no circumstances decide itself.

Three things that should be clarified before the first agent

  • Permissions: Which account does the agent work with, and what does this account have access to? Access to the mailbox is, in many companies, effectively access to contracts, HR matters and customer data.
  • Control points: Which actions run independently, which require approval? Drafting a reply email is something different from sending it to a customer.
  • Traceability: Can it be determined afterwards which action was triggered by whom or by which system? Without a log, every troubleshooting exercise becomes a guessing game.

Data protection is not an afterthought

The fact that Gemini Spark is available in the US first and Germany follows later is a familiar pattern. For companies, the delay means above all one thing: time to clarify the framework conditions instead of catching up on them later. Anyone who gives an agent access to business mail is processing personal data – the question of legal basis, data processing agreements and data flows arises regardless of how well the model writes.

In addition, agents read content that third parties have written. An email from outside can contain instructions that an agent misinterprets as a command. Anyone granting action rights should at least have considered this case before going live.

A pragmatic start instead of a fundamental decision

In practice, there is little to be said for handing entire processes over to agents straight away. More sensible is a narrow, well-defined use case with manageable potential for damage: summaries of long email threads, preparing draft replies, sorting incoming enquiries. Only once reliable results have emerged there over a few weeks does the step to actions with external effect become worthwhile.

Equally important is the question of who in the company maintains the task definitions. Agents are not software that is set up once and then forgotten – they need someone who checks whether the results still match reality.

What this means for web and software projects

In future, systems must not only be operable by people but also by software that acts on behalf of people: with clean interfaces, fine-grained permissions and logs that assign every action to an identity. Anyone planning a CMS, a portal or a specialised application today should treat these requirements as part of the architecture and not as a later add-on. The actual effort here rarely lies in connecting a model, but in clarifying responsibilities beforehand.

Sources