On October 9, 2026, Fred Cavazza publishes a French-language text on X devoted to an announcement he considers to have gone almost unnoticed: the shift from AI agents to synthetic collaborators. At its Gemini at Work event, Google presented a new generation of agents capable of handling complex missions, working in the background, and collaborating with humans or other agents. The post links to the Google Cloud blog post.
To situate this development, the author proposes four functional modalities of AI, of increasing sophistication. Chatbots answer questions, search, generate, or manipulate content, through one-off interactions. Agents execute tasks by mobilizing tools and applications and chain operations together, with variable autonomy. Orchestrators break down a complex need into an action plan, distribute work among specialized agents, and coordinate and consolidate results. Synthetic collaborators are persistent super-orchestrators equipped with their own digital identity (email, calendar, storage, memory); they receive missions from multiple people, work over the long term, and integrate into teams.
The author highlights three aspects of the announcement. The first is the unification of use cases within a single work environment that preserves context. The second is the organizational identity of agents: accounts, permissions, and resources make them identifiable actors within the information system. The third is the separation between the agentic harness and the model: Gemini also mobilizes third-party models, notably Claude, and memory, skills, tools, connectors, and orchestration together form a whole of which the language model is only one part.
Google would not be alone. According to the author, Microsoft is pursuing comparable ambitions with Copilot Autopilot, while Claude Cowork and ChatGPT Work remain limited to orchestration. These positionings are not sourced in the post.
Cavazza sees in this convergence among vendors a shift from an AI one queries to obtain answers to an AI to which one delegates operational responsibilities. Competition would no longer rest solely on model performance, but on the quality of semantic and methodological harnesses and on their integration into information systems. For companies, the challenge becomes organizing, governing, and supervising the work of these synthetic collaborators.
The text concludes with a caveat: the reliability and profitability of these collaborators at scale remain to be demonstrated, and the most advanced capabilities announced by Google are still in restricted access.