ChatGPT agents
Browser tasks, scheduled tasks and custom GPTs. Strong when the work lives in web apps.
- Home
- AI agents
AI agents, without the hype
AI agents that do real work. A person still signs off.
ChatGPT agents, Claude Cowork, Grok Bot, Meta’s Muse, OpenClaw. Everyone is testing them, and most firms end up with a cool demo and nothing running on Monday. I set up agents inside the tools you already use, give each one a single job, and keep a person on every approval.
Inbox agentRun in progress
- Wakes up on scheduleWeekdays, 7:00 am
- Reads new client emailOutlook, your account
- Sorts: reply, FYI, archiveRules you approved
- Drafts replies in your voiceNothing sent yet
- Waits for your approvalYou stay the gate
- Sends and logs what it didFull history kept
Illustrative run, modeled on a real client setup.
Plain English
Chatbot vs. agent.
A chatbot
- Answers when you ask
- You copy and paste the result somewhere
- Starts from zero every time
- Does nothing while you sleep
An agent
- Has one job and a schedule
- Uses the tools you gave it access to
- Does the steps, start to finish
- Stops where a person must decide
The landscape, translated
Which agent is which?
My job is picking the one that fits your firm, not the one trending this week.
Claude Cowork
A desktop agent that works with your files, folders and Microsoft 365. Strong for document-heavy work.
Microsoft Copilot
Inside Outlook, Teams and Excel. The natural fit when your firm already runs on Microsoft 365.
Gemini
Inside Gmail, Drive and Sheets. The natural fit for Google Workspace firms.
Grok Bot
xAI’s cloud agent, still in early beta. Built for long jobs that run unattended.
Meta Muse
A consumer personal agent for bookings, forms and shopping. Fun to try; not where client files belong.
OpenClaw
An open-source agent you host yourself. Maximum control, and the security is entirely on you.
Landscape as of October 2026. These products change monthly; the fit questions do not.
Real setups
Agents I’ve actually put to work.
01
Inbox triage agent
Checks Outlook on a schedule, sorts what came in, and drafts replies for approval. Nothing goes out until the owner says so.
“Okay. I love it. I'm going to have a clean inbox here by the end of next Friday.”
02
Report-to-client-email agent
Pulls a report from a business app in the browser, turns it into a visual, and drafts the client email in Outlook for review.
“Oh, I love it. Okay. This is great. Oh my goodness. I'm so excited.”
03
“Grill Me” thinking partner
A reusable skill that interviews you about a problem before the AI acts, so the plan is yours, not a guess.
“I just learned how to use the "Grill Me" prompt. Amazing 🔥 Immediately turned into insight which made us create a new system for taking action. Proves the point that we are only using 10% or its true capabilities. We havent even started yet 💪 Great work man.”
04
Review-queue matching
Routine payment matching moves on its own; anything unclear waits in a queue for a person to decide.
“I love that, because that was the pain that was going on.”
Guardrails
Agent rules I never skip.
One job per agent
A clear task and a clear definition of done. No do-everything bots.
Your accounts, not mine
Agents run on firm-owned logins and tools you already pay for.
Read before it writes
New agents start read-only, then earn the right to draft.
A person approves what leaves
Client emails, payments and filings wait for a human.
Every run is logged
You can see what it did, when, and why.
One switch turns it off
You can pause any agent without calling me.
Agent-fit check
Is it a job for an agent?
Pick one task. Answer five questions. Takes 20 seconds.
- Does this task happen at least every week?
- Do the inputs already arrive digitally (email, PDF, portal, spreadsheet)?
- Can you describe “done” in one sentence?
- Would a mistake be caught before a client sees it?
- Does it follow roughly the same steps most of the time?
0 / 5
Answer the questions to see the verdict.
Five yes answers is a strong agent candidate. Two or fewer means fix the process first.
Pick my first agent with AntonioHow to start
Decide, learn, or have it built.
“You're living in the future, dude.” · Felipe Carrasco
Pick the first agent
$750 AI Opportunity Review
90 minutes on one workflow: whether an agent fits, which tool, and the guardrails, in a written memo.
Set it up with me
$1,000 per month
Side by Side coaching: build and tune your own agents in private sessions, with Slack between them.
Have it built
$4,500 starting at
AI Systems Build: I build the agent in your accounts, test it with your team, and hand you the keys.
Questions
AI agents, answered.
What is an AI agent, in plain English?
A chatbot answers when you ask. An agent has a job, a schedule and access to specific tools, so it can do the steps for you, like reading new email, drafting replies or matching payments, and then stop where a person needs to decide.
Is it safe to let an agent touch client data?
It can be, with rules: firm-owned accounts, the smallest access that does the job, read-only first, a human approval before anything leaves the firm, and a log of every run. That is how I set them up. Consumer agents built for shopping and bookings are not where client files belong.
Will an agent replace my staff?
No. It takes the repeated steps: sorting, drafting, matching, reminding. Your people keep the judgment, the exceptions and the client relationships, and get hours back for that work.
Which is best: ChatGPT, Claude, Copilot, Gemini, Grok or Muse?
The one that fits where your work already lives. Microsoft 365 firms usually start with Copilot or Claude; Google Workspace firms with Gemini; web-app-heavy work suits ChatGPT agents. The recommendation is about your firm, not the tool of the week.
What does it cost?
Picking the first agent and its guardrails is part of the $750 AI Opportunity Review. If you want to set agents up yourself with guidance, Side by Side coaching starts at $1,000 per month. If you want it built for you, AI Systems Builds start at $4,500. AI tool subscriptions are separate and billed by the vendor.
Can AI agents use this website?
Yes. This site publishes its services as tools for in-browser AI agents (WebMCP), so an assistant can look up prices, estimate handoff costs or check whether a task fits an agent, with your permission.