Write the errand in plain words — the agent googles, calculates, and hands you a table

Here's the idea in one line: you type an errand in plain words — "compare three robot vacuums under $300 and build a pros-and-cons table" — and the app hands you back not ten links, but a finished table. Someone went to the internet for you, checked the prices, wrote down the downsides, and pulled it all onto one screen. That someone is an agent.

And here's what's new. A month ago, "building an agent" meant writing the loop by hand: the model thinks → you call search → you feed the result back → it thinks more → you check → repeat. That loop is the whole agent, and wiring it up was a week of fiddling. In May, at Google I/O, Managed Agents landed in the Gemini API: you make one call, and the server runs the whole loop for you — it spins up a full Linux machine in a sandbox where the agent googles live, runs code, saves files, and keeps going until the task is done. The docs put it plainly: one call, and the result comes back to you.

Why this one

Everyone has a boring "go look it up and pull it together" task. Pick a vacuum. Work out what a three-day trip would cost. Make a list of clubs near home. By hand it's half an hour of tabs and copy-paste. "Type it in words, get a finished answer" kills all that fuss. You'll actually use this yourself.

And there's less "magic" here than it looks. The hard part — that tool loop — now lives on the server. Your app is a pipe: send the errand, wait, show the result.

What you'll learn

• What an agent actually is. Not wizardry, but a loop: think → go to a tool → check → repeat. You'll watch it run live — and stop being scared of it.
• Why an agent needs tools. Search and running code are tool use. The model decides on its own when to hit the web and when to calculate. You'll hand it those arms and watch it use them.
• One call instead of orchestration. A year ago you'd write that loop yourself. Now you hand over the task and wait. You'll feel how much that lowers the bar.

A ready starter prompt

Don't ask for "an agent that can do everything" — it'll sprawl. Give it a role, an output format, and limits:

A strong prompt leaves no room for guessing: the agent's role is clear, it's clear it must go to the web, the exact table format is set, and so is the ban on making things up. The first result lands much closer to what you wanted.

What you end up with

You type into the box: "what would three days in Lisbon for two cost in July — lodging, food, flights from London." You hit send. A couple of minutes — the agent went out, checked lodging and flight prices, ballparked food, and a table appears under the box: travel, nights, food, and a total at the bottom. Not "here are ten links," but a put-together estimate. You never opened a single tab.

Start with one box and one run — and you'll have a thing that does in a couple of minutes what used to take an evening of tabs.

Source: Managed Agents in the Gemini API: one call spins up the sandbox and runs the whole agent loop (Google AI for Developers)