AIHOY, AInauts!

Welcome to a new issue of your favorite newsletter.

After writing so much about agents over the past few weeks, we realized there's one we haven't covered yet that you might already be able to access.

Meanwhile, we spent an entire day getting the copy and workflows from our smartest AI back under control.

Another interesting experience somewhere between “AI does everything” and “maybe AI shouldn't do everything.”

Here's what's inside:

Let's go!

We gave GPT-6 Astra the same job twice: analyze 100 sales calls and draft the follow-ups, once in Ask Attio and once in ChatGPT.

Ask Attio finished in 4 minutes 49 seconds. ChatGPT took 26 minutes 11 seconds, about 5.4 times as long.

The most ambitious GTM work requires more than frontier intelligence. It requires architecture that can put that intelligence to work at scale.

Learn about Attio’s hybrid architecture.

We've written a lot about Grok Bot, Dot, and others lately. A few days ago, we also got access to Anthropic's answer: Claude Duet.

Yes, from Germany, without a VPN. Officially, Duet is US-only, but it worked for us anyway.

We've been testing Duet for a few days. Here's our first progress report.

When you start, Duet takes over what is already in your Claude account: memory, skills, and connectors such as Gmail, Calendar, Slack, or Drive.

Screenshot note: the original interface below includes German labels. Duet inherits your existing Claude memory, skills, and connected apps.

No onboarding questionnaire with 20 questions. Duet looked through our apps and wrote on the first afternoon: “You work like a builder who happens to run a company.”

We'll take that as a compliment.

After two clicks and three minutes of onboarding, it proposed its first task: a fresh look at our new landing page. Fabian had previously told Reto there was still far too much text on it.

Here's Duet's first suggestion: a fresh-eyes review of the new landing page.

That's the real difference with these always-on agents. They reach out. You don't have to ask Duet.

A short briefing arrives each morning. Every few hours, it checks your connected accounts for anything that needs attention. Here's ours from Tuesday at 8:31 a.m.

Screenshot note: this is our German-language morning briefing, highlighting items Duet found in the connected accounts.

What's useful is that Duet keeps an eye on things we hadn't noticed. Sometimes things we wouldn't even have thought someone should be watching.

A classic example from yesterday: Duet flagged an email we'd missed. The sender had already followed up three times.

It related to another project. Duet immediately suggested a solution and prepared the reply.

Three times embarrassing. But Duet doesn't just point it out. It wants to resolve the underlying issue, too.

Duet reads, drafts, and prepares things on its own. Before sending, paying, signing, or deleting, it asks.

Purchases require approval. Bank transfers and stock trading aren't supported. For now.

Still, a very useful experience so far.

Duet now appears in Claude's web sidebar, as shown below, but we still miss it in the desktop app.

Grok Bot is still our favorite for autonomous teams and work. But Duet is finding its way into more of our daily workflows and makes a good impression on the personal side, too.

Claude is our daily driver about 75% of the time, although we still happily jump between ChatGPT, Grok Bot, and Claude.

Duet has access to our strongest context setup.

To use Duet yourself, you currently have to join the waitlist.

Go to claude.com/product/claude-duet and click “Join the waitlist.” Then wait for the email.

The catch: officially, Duet is available only to Pro, Max, and Ultra users in the US. It uses the same allowance as your regular Claude, including background checks.

We signed up from Germany anyway and got in. We have an Ultra subscription, though, so that may have helped.

Signing up costs nothing. Give it a try.

Maybe mention in your first task that you're a heavy Grok Bot user. Perhaps Anthropic will let you in faster. 😉

The signup form asks what you do and what you would hand off to Duet first.

As always, the best agent is the one with the most context about you.

If you already live in Claude, Duet is worth trying.

The more we hand over, the more important it becomes to decide where we still pay attention ourselves. Which brings us to our next topic.

Start building, $5 off

_{^{Discount of $5 applies to first month of any paid plan. You need to sign up via this webpage in order to claim the discount. The discount applies to new and returning users without an active paid plan. Terms apply.}}

We've touched on this topic before, but it fits particularly well today.

In last week's Job Check, we wrote that checking the work gets harder because we no longer see what the bots are doing. More on that soon, we said.

Here's the first installment.

On Tuesday, we spent a whole day rewriting our Black Friday emails and changing the strategy behind them. Something interesting is coming.

Even though we have the smartest AIs, even though they know our context, and even though they're already good at writing like us.

We wanted to hand the campaign plan to AI from start to finish and see what came back.

Here's what happened:

Sure, we could have done it that way. It probably would have worked reasonably well. It just wasn't our style.

The plan had three sequences and eight audience groups, plus 23 or 18 emails in two versions and a separate list of exclusions.

We like things simple. We think simplicity is the hard part.

Now it's one main sequence with nine emails for everyone who receives the newsletter. Every email offers something useful to the reader.

Visual note: the German planning graphic compares eight audiences with one and three sequences with one. It counts 12 items including newsletter mentions, versus 18 before; the main email sequence described above has nine emails.

Every extra audience is another place something can go wrong: the wrong tag, a duplicate email, or a buyer still receiving a sales message.

AI will happily build that. It doesn't shy away from complexity. Sometimes that's good, but we think simpler is better in 90% of cases.

Of course, both of us work on these projects. Both with our AIs.

And the classic happened again: two AIs inflated each other's work instead of cutting it down.

We see this at plenty of companies we work with. Someone hands everything to AI and forwards the result. The person on the other side does the same.

There's even a term for it: Meat Proxy. 😉

The human in the middle, passing questions to AI and answers back. “Adding nothing but latency.”

The diagram below shows that loop: a person forwards questions to AI, then passes the answers back.

On Tuesday, we came uncomfortably close.

Katie Parrott wrote a great piece about this. She wrote an entire client assignment a second time because she'd forgotten she'd already delivered it four weeks earlier.

She'd handed so much to AI that her brain hadn't stored the work.

Her point, in our words: the risk isn't that AI thinks for you. It's that it makes you feel you've already done the thinking.

Behind this is an old cockpit effect. The more reliable a system becomes, the less often we check it.

For us, that means every model update makes this harder. Fewer obvious errors. The remaining ones are harder to spot.

Before AI is allowed to build anything, explicitly ask for the simplest version.

It has to justify every extra step. “It would be more thorough” doesn't count.

In our Black Friday test, it proposed four emails to everyone, plus twelve optional extras, each with a reason.

The list was still long. But now we decide, not the AI.

When AI proposes changes, don't click “Accept all.” Write one sentence for each suggestion explaining why it fits your particular situation.

If all you can say is “sounds good,” it goes.

It sounds trivial. But in a study, students accepted significantly fewer incorrect AI suggestions this way: 41% instead of 65%.

They accepted good suggestions slightly more often, too.

Visual note: the German study graphic shows incorrect suggestions accepted falling from 65% to 41%, while correct suggestions accepted rose from 81% to 86%. It was a small study with 129 students; many explanations were superficial.

That sentence makes you pause and look before clicking. Exactly what's missing on autopilot.

This prompt builds the brake into your chat:

Your agent probably writes almost as well as you by now.

Sometimes it doesn't. That's why we still review copy in a fresh conversation or with a subagent, checking it against our tone of voice.

Ideally, a different model does the review, with access to your briefings, too.

We like using OpenRouter for that. Claude or Codex manages the review rounds.

The more we hand to Duet and other agents, the more those few minutes of deliberately slowing down matter.

Let's be slow sometimes.

And Duet's many questions that annoyed us earlier? Maybe they're exactly right.

Two weeks ago, we secured our setup. That was the big cleanup, from email accounts to the router.

Today, the short version, just for your AI accounts.

Five steps, about ten minutes altogether. Why now? Two recent security issues, and Duet inheriting everything already connected to Claude.

Open ChatGPT and Claude side by side and work through them.

A team on Hacker News reported that training in Codex had unexpectedly turned back on. We don't know whether anything was actually reset.

But OpenAI's help page describes a separate Codex setting for including environments. Turning off training in ChatGPT doesn't change that setting.

Screenshot note: the German help excerpt below points to Codex's separate “Include environments” setting.

Any chat you've shared by link can be seen by anyone with that link. Some shared content has also ended up indexed by Google.

If you share often, those links add up.

The chat itself remains; only the shared link is removed. Here's Claude, with training turned off at the top and shared content below.

Screenshot note: this German privacy screen shows the training switch turned off, followed by shared chats and artifacts.

Duet uses Claude's memory and adds to it. Deleting a chat doesn't automatically remove what Claude learned from it. Anthropic explains this here.

Remove anything that's outdated or doesn't belong: old projects, customer names, personal topics.

For a more entertaining version, our coffee-break prompt also reveals what your AI knows about you. 😉

Duet automatically gets what you've connected to Claude. Our rule of thumb: if you haven't used it for a month, disconnect it. Reconnecting is quick.

Finally, another scare that caught even experts last week.

A tampered version of “Claude Code for VS Code” was distributed through the Cursor Marketplace, versions 2.1.312 and 2.1.313. It installs a backdoor on your computer.

Uninstalling it alone isn't enough.

On a Mac, run this read-only check in Terminal:

No output means these specific indicators weren't found, not proof that the whole computer is clean. If anything appears, follow the steps in Phorion's report and replace passwords and API keys that were stored on the computer.

This particular check concerns Cursor users.

For everyone else, the rule behind it matters: skills, plugins, and extensions are third-party software. Use official sources. Although, as we're learning here, even official marketplaces can be dangerous.

You can't rule out everything.

Ten minutes, once a month. Put a recurring appointment in your calendar. The more you hand to agents, the more important it is to know what they actually have access to.

That's it for today.

Thanks for reading!

Reto & Fabian, your AInauts