Big vendors productize supervised agents
Microsoft and Google are testing always‑on, agentic assistants that act like supervised coworkers — and Google explicitly documents 'subagents' so a main agent can delegate specialist tasks without bloating context. GitLab is also integrating agent orchestration with Vertex AI and compliance controls, showing vendors are packaging delegation plus governance. (cnet.com) (androidauthority.com) (geminicli.com) (itwire.com)
The biggest technology vendors are turning artificial intelligence agents into supervised workplace software, not just chatbots, with Microsoft, Google, and GitLab all adding delegation and control layers. (microsoft.com) An agent is software that can take steps on a user’s behalf, like gathering files, drafting a report, or updating a system after getting instructions. Microsoft said on March 9 that Wave 3 of Microsoft 365 Copilot adds “embedded agentic capabilities” and a preview of Copilot Cowork for multi-step work inside its productivity suite. (microsoft.com) Google is testing a similar shift inside Gemini Enterprise. Android Authority reported on April 15 that a new “Agent” tab under test includes sections for Goals, Agents, Connected apps, and Files, plus a Tasks dashboard aimed at longer-running workflows. (androidauthority.com) Google has also started documenting how one agent can hand work to another instead of stuffing every instruction into one giant prompt. The Gemini CLI documentation says “subagents” are specialized agents that handle tasks like codebase analysis or documentation lookup “without cluttering the main agent’s context or toolset.” (geminicli.com) That design solves a practical problem in agent software: the more one model has to remember at once, the more expensive and error-prone it can get. Google’s documentation says each subagent can carry its own prompt, persona, and tool access, so the main agent can delegate specialist work and keep its own context smaller. (geminicli.com) Vendors are also selling the management layer around those agents, not just the models underneath them. GitLab said on April 14 that its Duo Agent Platform can now call foundation models through Google Cloud Vertex AI, including Gemini models, while routing actions through GitLab compliance and audit controls. (about.gitlab.com) GitLab said customers can run its artificial intelligence gateway on Google Cloud without separate infrastructure and count Duo Agent Platform usage against existing Google Cloud commitments. That ties agent orchestration to procurement, security review, and audit logging, which are the parts large companies usually need before deploying automation widely. (about.gitlab.com) Google has already been packaging agents as a product inside Workspace. In December 2025, Google announced the general availability of Workspace Studio, which it described as a place to “design, manage, and share AI agents” across Google Workspace using Gemini 3. (workspace.google.com) Microsoft is framing the same shift as a move from assistance to execution. Its March 9 announcement said Copilot is moving “beyond assistance” and that Copilot Cowork can use multiple models so the system can choose the right model for a job rather than relying on one vendor’s model alone. (microsoft.com) The common thread is that “agentic” features are being packaged like software coworkers with supervisors, specialists, and logs. The next contest is less about who has a chatbot in a sidebar and more about which vendor can make delegated work reliable enough for everyday business systems. (microsoft.com)
