A few weeks ago, "AI agent" still sounded like a buzzword. This month, it got much more practical.
Google announced Gemini Spark, a personal agent that can plan tasks across connected apps. OpenAI's ChatGPT agent is available on paid plans and can browse websites, work with files, fill out forms, and use connected apps. Microsoft made its "computer use" agents generally available inside Copilot Studio.
That's a real shift. But there's a familiar problem hiding underneath. Most people are going to hear about agents, watch one impressive demo, and either try to automate something too big or never try it at all.
Here's the better path. Pick one small task, give an agent exactly one hour with it, and see what actually happens. This week is about how to do that without breaking anything.
🗞️ THIS WEEK IN AI
1. Google announced Gemini Spark, a personal AI agent
At Google I/O this month, Google announced Gemini Spark, a new agent for the Gemini app that is designed to work across Gmail, Docs, Slides, and other connected tools. Google says Spark can take multi-step actions under your direction, including background tasks.
Why this matters: This is a strong signal that personal agents are moving from demos into everyday work tools. But it is not broadly available yet. Google says Spark is rolling out to trusted testers first, with a beta planned for U.S. Google AI Ultra subscribers.
2. ChatGPT agent is available on paid ChatGPT plans
ChatGPT agent can navigate websites, work with uploaded files, connect to data sources like email and document repositories, fill out forms, and edit spreadsheets while pausing for clarification or confirmation when needed. OpenAI says it is available on Plus, Pro, Business, Enterprise, and Edu plans in supported countries.
Why this matters: This is one of the most accessible ways for regular users to try an agent today. It is not just a chatbot giving advice. It can actually move through a task, while still keeping you in control.
3. Microsoft's "computer use" agents became generally available in Copilot Studio
Microsoft made its computer-use agents generally available inside Copilot Studio. These agents can click through software interfaces the way a person would, which lets them work with older tools or internal systems that do not have a clean built-in connection.
Why this matters: If your work involves clicking through the same five screens every Monday morning, that's exactly the kind of task this is built for. Your IT team may already have access to it.
🛠️ THIS WEEK'S TRICK
Give an AI Agent a One-Hour Trial Run
The mistake most people make is picking a task that's too important for a first run. Either the agent fumbles it and you write off the technology, or it works and you start trusting it for things it shouldn't be trusted with yet. The fix is to pick a task that's small, repeatable, and easy to undo.
Step 1: Pick the right first task
You want something that hits all three of these:
Low stakes (nothing breaks if the output is bad)
Reversible (you can throw the result away with no consequence)
Boring (you'd be happy to never do it manually again)
Good candidates: draft replies to three specific emails, build a one-page weekly recap from your calendar, summarize a long PDF, pull action items out of a meeting transcript, or compare three vendor pages and turn the notes into a simple table.
Bad candidates for a first run: anything that touches money, sends external messages, or changes shared files your team relies on.
Step 2: Scope it with a prompt before you start
Open ChatGPT, Claude, or another regular chat assistant first. Run this prompt before you hand the work to an agent:
I want to try an AI agent on this task: [describe the task in one sentence].
Before I run it, help me sanity-check the scope:
1. Is this task small and bounded enough for an agent to finish in under an hour?
2. What inputs does it need from me up front?
3. What's a clear definition of "done" so I know whether it worked?
4. What's one thing that could go wrong, and how would I catch it?
Don't tell me how to do the task. Just help me set it up well.
Use the answers to tighten the task before you hand it to the agent. If the answers show the task is too big, shrink it.
Step 3: Watch the first run, don't walk away
This is the part most demos skip. Stay at your desk for the full hour. Watch what the agent does at each step. If it asks for confirmation, take a second to actually read what it's about to do before you click yes.
When it finishes, ask yourself two questions:
Did it match the definition of done from Step 2?
Would I trust it to do this again next week without me watching?
If yes to both, you have a real workflow. If no, you've learned something useful in an hour, which is a fine outcome.
What this works best for:
People who keep hearing about agents and feel behind
Repetitive tasks you've been meaning to automate for months
Anyone whose company "rolled out an AI tool" without showing them what to do with it
A note: most of the value in the first run is not the output. It's learning what an agent is actually good at versus what the marketing says it is. That changes how you'll use these tools for the next year.
🔧 TOOL OF THE WEEK
ChatGPT agent - chatgpt.com
ChatGPT agent is a mode inside ChatGPT that can work through online tasks instead of only telling you what to do. It can use a browser, work with uploaded files, connect to apps, fill out forms, and edit spreadsheets while asking for confirmation when needed.
For an issue about trying an agent on something small, this is the cleanest place to start for most readers. You do not need to set up a developer account, install automation software, or build a workflow from scratch.
Price: Agent mode is available on paid ChatGPT plans, including Plus, Pro, Business, Enterprise, and Edu. Free ChatGPT accounts do not include agent mode.
Best for: People who want to test an agent on research, form-filling, file work, spreadsheets, or web tasks without learning a new tool.
Worth trying? Yes, if you already have a paid ChatGPT plan. Keep the first task low-stakes and stay nearby while it works.
💡 PROMPT OF THE WEEK
Use this when you are deciding whether a task is even worth handing to an agent in the first place. Works in ChatGPT, Claude, or another regular chat assistant.
I'm thinking about delegating this task to an AI agent: [describe the task].
Help me decide if it's a good fit by answering:
1. What part of this task is rules-based and easy to automate?
2. What part requires judgment a person should still make?
3. If I split it that way, what's the smallest version I could test this week?
4. What signal would tell me it's working well enough to keep using?
Tip: if the answer to #2 is "most of it," that's a sign to keep doing the task yourself for now, and use AI as an assistant instead of an agent.
👋 THAT'S A WRAP
Agents are no longer just a thing on the horizon. Some are already usable, and others are clearly on the way. But the way to get good at them is not by trying to automate your whole job. It is by spending one hour on one task and actually watching what happens.
Pick a task this week that fits the three rules: low stakes, reversible, boring. Run the scoping prompt. Then give an agent a shot at it while you watch. You'll learn more in that hour than from any podcast or LinkedIn thread on what agents can do.
If you know someone who keeps saying they "need to figure out AI agents" but has not actually tried one, forward this to them. The barrier to a first try is lower than they think.
See you next Friday.