Last week at Google I/O I recorded a podcast with Josh Woodward — the VP at Google responsible for Gemini and Google Labs. Josh has been at the company for 16 years. He helped build Notebook LM, AI Studio, and Flow.

Here's one thing from our conversation I want to dig into today.

The reason: everyone on his team is managing agents now.

For the past 2 years, the big promise of AI was: do the same things faster. Write faster. Research faster. Summarize faster. That phase is ending.

I was using AI for individual tasks at first: write an email, make a summary, translate something. Then I started feeding it context — my best writing, my channel stats, my working principles. And eventually moved from individual tasks to setting goals. In the first case, I get a completed task. In the second, I get a tool that understands what I'm building and works toward that.

Josh calls it the shift from doing to directing — you are no longer the one executing the task. You define the outcome, evaluate the result, and decide if it's good enough. The skill required here isn't knowing how to code. It's the ability to think clearly about what outcome you want, to build a system where the tool knows your context, and to recognize when it's moving in the right direction versus drifting off course.

If you're already managing people, you've felt this for a while. Now everyone else is catching up. That's why Google is starting to think that way more people in the company might need a management training, because everyone is increasingly managing agents.

Most AI tools today run at 50–100 tokens/sec. Fast enough to feel real-time in a chat window. But a 10-step agent task — research something, draft a response, format it, send — takes several minutes at that speed. Annoying enough that you end up doing it yourself.

At 1,500 tokens/sec, that same chain runs in seconds.

Google has it running internally right now. The bottleneck that makes agent workflows feel unreliable today is a temporary problem. It's going away, and faster than most people expect.

For the past two years, working with AI meant learning to write better prompts — precise, structured, specific. That skill still matters. But voice is quietly becoming the more natural interface, because it removes the translation step between what you're thinking and what you type.

Google's announcement of Gemini Spark at I/O last week is a good marker of where this is heading: a voice mode that doesn't convert speech to text, but converts speech to action. You talk through what you need, the system reads your files, mails or calendar, finds the context, and returns a finished document.

This is not transcription. In regular transcription, voice gets converted to text. Here, voice becomes a task: you explain what you need, and the system goes into your files to find the right context itself.

Josh demonstrated live by saying a wrong date on purpose mid-explanation. The system caught it and fixed it without a follow-up prompt.

When AI executes faster than you can review, what matters most is judgment.

Can you tell in 60 seconds whether a 500-word draft has the right argument? Can you catch when an agent completed a task 85% correctly and articulate the 15% that's wrong? Can you write a brief specific enough that the output doesn't need 4 rounds of corrections?

That's not prompt engineering. It's taste. Editorial judgment, a clear internal standard.

That's what I'm actively developing in myself and in my team right now — and it's kind of the most interesting professional skill I've worked on in years.

The full interview with Josh is on the channel. The conversation goes deeper there — including a separate section on why human taste, in Josh's view, will become more valuable as AI takes on more of the work.

Watch the podcast on YouTube ↗

Marina 💜