The change lies less in the model’s name than in the nature of its intervention. A conventional assistant drafts text or summarises a document. Google’s agent is meant to break down an objective, select tools, connect to authorised systems and return work in an email, document or development environment. This remains a product announcement: real use must establish reliability, gains and supervision costs.

From an assistant that suggests to an agent that acts

Google says the agent plans work and selects a suitable model, currently from Gemini and Anthropic’s Claude family. Specialist versions for finance and legal work are in preview.

A presentation describes intended capabilities; it does not demonstrate behaviour at scale. Companies must separate promises from results measured in their own processes.

Permission becomes the central issue

Reading a shared file, preparing a reply and changing a contract do not carry the same risk. The agent should receive only the minimum rights required for a defined task and period.

Activity logs must show sources and proposed actions. Sensitive operations should remain subject to explicit human approval.

Tasks change before occupations disappear

Research, meeting preparation, synthesis and first drafts may be accelerated. In return, employees will spend more time checking outputs and reporting errors.

Success should be measured through actual time saved, human correction rates, cost and data protection—not simply the number of tasks delegated.

Conclusion

The announcement confirms the shift from AI that advises to AI that acts. Trust will depend on narrow permissions, auditable records and a permanent human right to stop, correct or reject an action.