This is the story of Agent Designer in Gemini Enterprise: a way for the people who know a job best to build the AI agent that does it, just by describing it.
So much work repeats. The weekly report, the inbox triage, the product email. It runs on a schedule, and the people doing it know it by heart.
People feel it. In Microsoft's 2023 survey of thirty-one thousand workers, six in ten said they spend too much of the day searching for information, and almost seven in ten said they don't get enough time to focus.
And the people who'd automate that work are in marketing, sales, HR and IT. Gartner forecast that by 2026 at least eighty percent of low-code users would sit outside IT. They're experts in their work, not in code, so building an agent had to feel like explaining the job to a new colleague.
There was no single number behind this. The case was the race to build agents, and the chance to let clients build their own, in ways they could understand and shape. Google didn't want to impose one way of working; it wanted people to make their own workflows. Some agents would come from the company, for everyone. Others, any individual contributor could make for themselves. So it was efficiency built as a community, and one person at a time. And the appetite was real: in PwC's survey, nearly four in five US executives said agents were already being adopted where they work.
These were very fast sprints. We built overnight, and on weekends too, and the PRDs took shape as we went. It meant endless conversations between engineering and design. Product managers and designers could vibe code by then, so suddenly we were making clickable, fully working prototypes. Then we'd hit a wall when engineering raised guardrails, privacy and permissions. There was hardly time for research: we were moving with the industry, and with what product leadership asked for.
You say what the agent should do, and it appears as a flow of steps you can read. If something's off, you don't rewire it. You just ask for the change.
Making it real meant the whole path: a canvas for the flow, the prompt and settings behind it, test runs to see it work, and publishing when it's ready.
The bigger idea: once a few teams have built agents, nobody should start from zero. The Agent Gallery lets people find one that already does the job and start from there.
What mattered most was launching simple agents. Sub-tasks and sub-agents would come later, and so would scheduling and evaluation. The gallery came second, on a hunch: that most people would use agents someone had already made, rather than build their own.
An agent out in the world needs someone watching. Admins can follow its runs, its errors and how much it's used over time.
Gemini Enterprise launched in October 2025 with a no-code workbench for building agents. By April 2026, Agent Designer and the Agent Gallery sat at the heart of it.
What I take to my students from this one: design for the people who will live with a future, not only for the ones who will pay for it. Thanks for watching.