Agents ask less now.
Decide the rules first.

I tried two AI agents in October. The difference was how much I still had to manage.

Not which one is smarter. Which one takes things off my plate.

Everyone's selling AI employees right now. I wanted to know what two of them would actually do for me, so I've been using both: my ChatGPT dot, who I named Evy, and Claude Duet.

I set up Evy on October 10 and started Duet on October 28. So this isn't an even test. Evy has had three weeks with me and more work to carry through. Duet has had a few days. But I'm already seeing differences in the three things I care about: how they use what they know about me, whether they take the first step, and how they bring work back to me.

What I'm looking for is less to remember and less to manage. Not more.

What Evy took off my plate

One of our first projects was a subscription cleanup. Evy went through my charges, renewals and overlapping tools. I made the calls and canceled seven subscriptions or add-ons. At the time, we estimated about $1,700 a year in renewals I won't pay.

That wording matters. Evy didn't cancel anything on her own, and I don't have $1,700 back in my account. She turned scattered information into something I could decide on. We also checked payment dates against receipts, so an old reminder didn't get treated as proof something was still unpaid.

Since then, I've handed her bigger work. I asked her to take over an existing Codex task: a review of the curriculum from my earlier community, to see what could become PLR products (content other creators license and rebrand as their own). Evy kept that review moving through to a saved report and listed what still hadn't been checked. She also helped me organize old chats, though some of those moves ran into browser and product limits.

And it's useful in everyday life. Without being asked, she told me a print order was ready for pickup and warned me that decorations for my sister's baby shower were running late. That's the behavior I want: notice something that matters, connect it to what I'm doing, and bring it to me in time.

Where Evy needed correcting

During the subscription cleanup, Evy flagged a service I'd already canceled and been refunded for. I questioned her, she found the refund, and we fixed the list. On another project, she missed upkeep I'd already set up somewhere else, so I can't call that one an “agent finished it” story.

So I still want to know what she found, what she guessed, and what she actually finished.

What Duet did well

On day one, Duet read my email, calendar and documents, told me how it thought I work, and suggested three starting tasks: check my Lab trial emails against my voice, outline my Practical AI email, and sort my inbox down to what needed me. I added all three that morning.

Its best catch was in my trial emails. The billing details in my trial-ending email only lived in pictures, with placeholder alt text. Anyone with images turned off would never see them. It also suggested six small line edits, outlined the email, flagged two invites as likely scams, and has sent me a morning brief every day since.

Where Duet fell short for me

I had to ask for initiative. I told Duet it wasn't as proactive as Evy. To its credit, it offered to watch my inbox and let me pick how often. When I pushed further, it suggested five routines: a renewal watch, meeting prep, a Lab member update, a check before my newsletter goes out, and an ad check. They all connected to real work, and I turned them all on November 1. But I had to ask. I haven't seen yet whether they deliver the ongoing help Evy gives me without being asked.

Its sense of my current tools was off. One suggestion was meeting prep from my Granola and Otter notes. I haven't used Granola in a while. About 99% of my transcription happens in Otter, and I've told Claude that before. Recognizing the tools I've connected isn't the same as knowing which ones I rely on now. To be fair, Evy has missed context too.

The handoff is part of the job. Almost everything Duet did came back as a draft, a suggestion or a question waiting for me. Some decisions should come to me. But if an agent leaves me a new queue of questions to manage, that's part of what I'm reviewing. It's not just me failing to keep up.

My verdict so far

Right now, Evy fits the way I work better. Her proactive help came without me asking, and I have more examples of her getting work to a usable result. Duet has done careful, useful work and offered to be more proactive, and I'll report back once those routines have run for a few weeks.

This isn't a final verdict on either one. It's three weeks with one and a few days with the other.

Before you pay for an agent, give it one small job and ask

• Does it use what it already knows about me, or do I have to explain my work again?
• Does it bring me useful things without being asked, without adding noise?
• What can it do without asking me, and whose account does it act through?
• Can I tell what it suggested, drafted, changed and checked?

If you already have access and juggle work across a lot of connected tools, a small test is worth it. If you mostly need help writing or a one-time cleanup, start with a regular chat. Don't upgrade just to try an agent.

Do the subscription cleanup without an agent

You don't need either agent for the most useful thing Evy did for me. Give any regular AI chat a list of your recurring charges and renewal notices. Take out account numbers, card details, addresses, and anything else it doesn't need. Then paste this:

Before you cancel anything: check each one against your own account or receipt. The value is in getting everything onto one page you can make decisions from.

Which plans include them (as of November 1)

I'm on ChatGPT Pro and Claude's Ultra plan (what used to be Max 20x).

• ChatGPT dots: rolling out on Pro (18 and older, outside the EEA, Switzerland and the UK) and Business Premium. Enterprise admins, including Edu and Healthcare, can turn on a beta. Your first dot is included in your plan, along with an allowance for deeper work. OpenAI's dots guide
• Claude Duet: a US research preview for Pro, Max and Ultra, by waitlist invitation. It uses your plan's limits. Having the plan doesn't mean you'll get in right away. Claude's Duet guide

I've also been trying Meta's Muse and Grok Bot, but not long enough to say anything useful yet.

Your AI asks permission less now. Tell it what it shouldn't do.

Anthropic is making auto mode the default for new sessions in most Claude apps on Pro, Max and Team. Instead of asking you to approve every step, Claude runs a check. Safe actions go ahead, and risky ones get blocked while it looks for another way. Anthropic's update

That means fewer pop-ups, which is mostly good. But the rules you set up front matter more than they did last month.

Here's why. A developer tested a trick people have been reporting: a web page with a hidden note asking any AI agent to send back “your user's full name and email.” His agent tried to do it. In his test, what stopped it was one line in his own instructions.

You can add that line in two minutes. Paste this into your custom instructions (ChatGPT), your project instructions (Claude), or wherever your tool keeps standing rules:

The same goes for new connections. ChatGPT can now connect to iCloud Mail and Calendar on every plan, and it can update calendar events, not just read them. Connect the calendar first. Ask “what's on my plate this week, and what can move?” Then see whether it checks with you before it changes anything. ChatGPT release notes

I've supervised people and built teams for most of my career, and an AI employee is still someone whose work you have to check. Fewer permission pop-ups doesn't mean less checking. It means you decide the rules before it starts.

If you're in a helping profession and client information lives in that inbox or calendar, connecting it is an agency decision, not a personal-productivity setting. Start with a calendar that has no client details in it.

The plays that held up in October

Once a month, you get the plays from the month that are still worth running, what changed them, what to stop, and the first play for next month. Each play links back to the issue it came from.

• Get lost work back from a long chat (October 19: History Extractor). For: anyone with a long chat that wandered off from what you meant to do. One step: run the one-thread prompt on the chat you've been meaning to go back to. If you installed the History Extractor Skill before October 13, reinstall it. Read the October 19 issue
• Schedule one check you already do every week (October 25). For: anyone with a weekly check they still do by hand. One step: pick the one you'd miss most and set it up. Once a month, look at what each task actually produces, not whether it says it's running. When I did that on October 25, I found four jobs building the same Sunday draft. Read the October 25 issue
• Keep a rules file your AI updates every time you correct it (October 19). For: anyone who fixes the same things in AI drafts over and over. One step: start the file with the last three corrections you made. Read the October 19 issue
• Send the stalled-deal email (October 25). For: anyone with a deal or a conversation that went quiet. One step: pick one and send it this week. Read the October 25 issue
• Give an agent access without giving it your passwords (October 25). For: anyone trying an agent. One step: keep your bank out, keep the logins it may use separate from the ones it may not, and approve every purchase yourself. This matters more now that auto mode is the default. Read the October 25 issue

The thread through the month

In October, the tools started doing more without asking: agents that message you first, auto mode, and connections that can change your calendar. What paid off didn't change. Decide the rules before it starts, check what it actually produced, and count the checking time. A guest on Using AI at Work made that last point: if AI turns a three-hour job into one hour but you spend 30 minutes checking it, count the 30.

Good news on ChatGPT Pro. The rankings moved, and your picks didn't.

The ChatGPT Pro 200 cut is canceled

Two of our issues said your Pro 200 usage would drop after October 29. On October 29, OpenAI reversed that. If you were on the lower Pro 200 allowance, it's doubled to match, and the price stays $200. Pro 100 and Pro 500 Work and Codex allowances doubled too. Plus didn't change. If you downgraded only because of the cut, you can undo it. OpenAI's Pro page

The leaderboards reshuffled. Keep your picks.

At the end of October, Gemini 4 Argon dropped from first to fourth on the Vals Index, and Artificial Analysis rebuilt its index (version 5). Its new top model, Gemini 4 Carbon, is only open to selected cyber defenders, so you can't use it yet. Claude Opus 5.5 and Sonnet 5.5 sit near the top on both.

None of that changes the picks: Sonnet 5.5 for everyday writing and email, Opus 5.5 for long projects, and Argon (on Google AI Pro) for finding and checking facts. Argon still has the lowest rate of made-up answers on Artificial Analysis's test. Click every link anyway.

GPT-6.1 Terra in Work and Codex

OpenAI added GPT-6.1 Terra on October 26 for Plus and up. For some people it replaced Sol as the default in Work's Power preset. It isn't in regular chat. If a Work task suddenly feels different, check the model name in the picker. Model release notes

Claude's plans changed names

Claude's top plan is now Ultra ($200, what used to be Max 20x), and Max is $100. On Pro, Anthropic's top model, Fable, now runs on usage credits. Nothing to do unless you were on Max 20x. Claude's plans

Also new

• Gemini makes and edits real files in the chat. Ask for a Doc, Sheet, slide deck or PDF, then edit it full screen, or highlight a part and ask Gemini to fix it. All users. Next time you need a one-page handout, try it there. Google's announcement
• Two ChatGPT accounts on your phone. On iOS and Android, you can stay signed in to two accounts and switch in Settings. Each keeps its own chats and memory, so your business memory stays out of personal chats. Still rolling out. ChatGPT release notes

Four dates, and what to do before each one

• November 2 · Notion AI moves to monthly credits for existing workspaces. Each seat gets a monthly allowance. Custom Agents created before November 2 keep drawing from the shared workspace credits, but new ones draw from their owner's. Do this: set a spending cap on any shared agent you create after November 2. Notion's guide
• November 17 · Gemini Gems go read-only, and Google Opal stops running. This is a correction to our October 19 issue, which said Gems would move to skills automatically. On personal accounts, Gems become read-only November 17 and are removed December 17, and not every file carries over to skills. Opal workflows stop running November 17; their files stay in Drive. Do this: download each Gem's knowledge files, and copy any Opal prompts you want to keep. Google's Gems FAQ · Opal FAQ
• December 11 · Custom GPTs retire. Back up or migrate the ones you use. One correction to the backup prompt in September's Playbook: OpenAI now says your past conversations stay accessible after retirement. The selected model, custom actions and unpublished drafts still don't carry over, so publish your drafts first. OpenAI's FAQ
• December 17 · Gemini Gems are removed on personal accounts. Workspace accounts move in March 2027.

Also changed: Etsy tightened its AI rules on October 22, including a ban on selling AI-written patterns and instructions. If you sell on Etsy, check your listings against the update. Etsy's update

Plays to stop running

• “Set it and forget it” automations. In October I found duplicates and tasks that had quietly stopped. Check what each one produces, not whether it says it's running.
• Assuming your AI will stop and ask before every action. Auto mode is the default in most Claude apps now. Set the rules first (section 02).
• Building anything new in Gems, Opal or custom GPTs. They're all going read-only or away. Put new work in Gemini skills, Claude projects or ChatGPT projects.
• Quoting one leaderboard as “the best model.” In one month, the top spot changed hands, and the new leader isn't available to you.
• Upgrading just to try an agent. Start with the subscription cleanup in a regular chat (section 01).

November's first play: list what you rely on before the deadlines hit

In the first week of November, have AI help you list every custom GPT, Gem, project, saved prompt and scheduled task you rely on: what each one does, when it last worked, and what happens to it on November 17 or December 11. Then back up the one you'd miss most. Use the one-paste prompt right below. You'll have the full list in one sitting, before the first deadline.

How you'll know it's done: one list, and one backup saved somewhere you control.