Latest / AI Fire Daily / #124 Robin: Claude Duet Wants to Become Your Always-On AI Assistant - But There’s One Big Catch

Transcript

#124 Robin: Claude Duet Wants to Become Your Always-On AI Assistant - But There’s One Big Catch · AI Fire Daily

• 0:00
  Imagine opening a coffee shop in 26 days. Stressful.
• 0:04
  Very. Your laptop is finally closed. You're trying
• 0:06
  to sleep. But down in the digital background,
• 0:09
  your AI notices you forgot to file a health permit.
• 0:13
  It quietly drafts it for you. It cross -references
• 0:16
  your calendar, automatically reschedules a less
• 0:18
  urgent task, and leaves an alert for your approval.
• 0:20
  It's working while you rest. Welcome to the Deep
• 0:23
  Dive. Today we're exploring Anthropix's newly
• 0:25
  announced Claude Duet. It's a massive shift.
• 0:28
  We are moving from reactive tools to proactive
• 0:31
  background workers. It's a total paradigm shift,
• 0:33
  yeah. Today we'll unpack what it is and how it
• 0:36
  functions offline. We'll look at the vital safety
• 0:39
  guardrails and we'll discuss the one major catch
• 0:41
  regarding usage limits, especially in this ongoing
• 0:44
  AI race against Grokbot and others. The landscape
• 0:47
  is definitely crowding fast. Okay, let's unpack
• 0:50
  this. Claude DeWitt has built for busy professionals
• 0:52
  and small business owners. Right. Which is very
• 0:54
  different from Claude Code. purely for developers.
• 0:58
  Exactly. Developer workflows are highly structured.
• 1:01
  A small business owner's workflow is completely
• 1:03
  chaotic. It's spread across dozens of platforms.
• 1:05
  I still wrestle with keeping track of all my
• 1:07
  open tabs and deadlines myself. It's exhausting.
• 1:09
  We all do. To really grasp this, we should define
• 1:12
  what an AI agent is. An AI agent is a digital
• 1:15
  assistant that takes actions for you, not just
• 1:18
  chats. That's the key distinction. Chat bots
• 1:21
  sit there waiting. Agents take the initiative.
• 1:24
  Right now, Dewate is available as a standalone
• 1:26
  app and on Claude .ai. It is a US -only research
• 1:30
  preview. And you do have to join a wait list.
• 1:34
  Yeah, but it's included in the Claude Pro and
• 1:36
  Max plans. And it's powered by Claude Opus 5
• 1:39
  .5. Which is their heavy hitter. Yeah, it handles
• 1:42
  complex multi -step logic. So it sends this morning
• 1:45
  briefing of important tasks. It scans your apps,
• 1:48
  checks updates all day, and prioritizes urgent
• 1:51
  work. Like spotting that missing health permit.
• 1:53
  It knows that permit is way more urgent than,
• 1:55
  say, a client renewal will do in three weeks.
• 1:58
  It's like a diligent chief of staff. And it can
• 2:01
  tackle bigger projects, too. Oh, absolutely.
• 2:03
  Dashboards, sales forecasts. It's trying to offload
• 2:07
  your mental project management. But I'm curious
• 2:09
  about the execution here. Does it just send these
• 2:12
  permits off on its own? No. It queues up the
• 2:15
  priority work, explains why, and waits for your
• 2:17
  final approval. So it queues up the work but
• 2:19
  waits for your green light. Exactly. You are
• 2:21
  the safety valve. That's reassuring. But if it's
• 2:24
  teeing up work across all our apps, how does
• 2:26
  it connect those moving parts without becoming
• 2:28
  chaotic? Well, Anthropic designed it to just
• 2:31
  slide right into your workflow. It carries over
• 2:33
  your existing cloud memory, your saved skills,
• 2:36
  and your connected apps. Okay, so no blank slate.
• 2:39
  Right. But here's the wild part. You use one
• 2:42
  continuous conversation for everything. Voice
• 2:44
  or text, it's all one thread. Unlike Grokbot's
• 2:47
  multi -agent approach. Yeah, Grokbot uses specialized
• 2:50
  agents. Duet just forces everything into one
• 2:53
  ongoing chat. Which seems like a limitation.
• 2:55
  It is a massive gamble. Mixing payroll data with
• 3:00
  marketing copy in one thread could confuse the
• 3:02
  AI's attention mechanism. We'll have to see how
• 3:04
  that centralized design holds up. But setting
• 3:06
  the interface aside, here is what really stopped
• 3:09
  me. The cloud browser? Yes. Imagine it grinding
• 3:13
  away on a server while your phone is off. It
• 3:16
  separates compute from your hardware entirely.
• 3:18
  It just wakes up every few hours, checks your
• 3:20
  accounts, and goes back to sleep. It saves compute
• 3:22
  but keeps a persistent pulse on your business.
• 3:25
  And it connects with over 6 ,000 apps. That's
• 3:27
  where the utility is. It's like stacking Lego
• 3:29
  blocks of data. You plug in QuickBooks for unpaid
• 3:31
  invoices and then Stripe to draft the reminders.
• 3:35
  You connect Gustover Payroll. You connect DocuSign
• 3:38
  for contracts. But again, it needs approval for
• 3:41
  emails, signing docs, or deleting things. And
• 3:44
  there are hard restrictions. It absolutely cannot
• 3:46
  trade stocks or move money. Right. It is mathematically
• 3:49
  blocked from doing that. A good firewall to have.
• 3:53
  But regarding that interface, are there separate
• 3:55
  threads for my payroll and my marketing? Not
• 3:58
  currently. Duet forces everything into a single
• 4:01
  ongoing thread, which is a potential downside
• 4:03
  compared to rivals. Got it. One mega thread instead
• 4:05
  of organized specialized folders. Yeah. It takes
• 4:08
  a mental shift to trust that. Handing an AI the
• 4:10
  keys to QuickBooks and DocuSign in one megathread
• 4:13
  sounds incredibly risky. How is Anthropic securing
• 4:17
  this? That is the big tension right now. Utility
• 4:19
  versus data security. Here's where it gets really
• 4:22
  interesting. Duet never actually sees your passwords.
• 4:25
  Never. It uses built -in checks to it won't share
• 4:28
  credit cards or government IDs. And health or
• 4:31
  financial data is strictly opt -in. If we connect
• 4:33
  this to the bigger picture, they're using OAuth.
• 4:36
  Secure login tokens. Yeah, it was tested heavily
• 4:38
  by internal red teams and an outside security
• 4:41
  firm. They really tried to break it. And early
• 4:43
  testers seemed to love it. The creator of Clodcode,
• 4:46
  Boris Cherny, uses it on walks. Just talks through
• 4:49
  his to -do list, entirely untethered from a desk.
• 4:52
  There's also a real estate broker whose AI cut
• 4:54
  a property showing schedule for the wrong day
• 4:57
  totally automatically. She didn't even prompt
• 4:59
  it. And a restaurant owner with 110 employees.
• 5:02
  They use one agent for accounting, support, and
• 5:04
  marketing. It acts as a unified operational layer.
• 5:07
  But I need to be clear on this. How does it access
• 5:10
  QuickBooks if it can't see my password? It uses
• 5:13
  secure login tokens in the background, so the
• 5:16
  raw password is never exposed to the agent. It
• 5:19
  gets a digital VIP pass without knowing the actual
• 5:22
  secret password. Precisely. And you can revoke
• 5:25
  that pass instantly. It's smart, it's secure,
• 5:28
  and it catches schedule mistakes. But there is
• 5:31
  a hidden cost to all this background magic. There's
• 5:33
  always a hidden cost. The biggest catch is the
• 5:36
  token economy. Duet shares your existing Claude
• 5:40
  Pro and Max usage limits. Right, and every offline
• 5:43
  background check -in burns tokens. Demanding
• 5:45
  casks like financial models burn even more. If
• 5:48
  it's waking up constantly, you'll hit your limits
• 5:50
  fast. Opus 5 .5 is efficient, but frequent checks
• 5:53
  really add up. The AI rivalry is exploding right
• 5:57
  now, which makes these limits so important. The
• 6:00
  timeline is insane. Grokbot launched in August
• 6:03
  2026. Metamuse and chat GPT dots hit in September.
• 6:07
  And now Claude Duet in October. Four massive
• 6:09
  players in three months. And their approaches
• 6:12
  are so different. Grokbot plans a multi -model
• 6:15
  approach. It chooses the best model for the task,
• 6:18
  even competitor's models. Yeah, an agnostic router.
• 6:21
  While Duet just uses Anthropix Opus 5 .5 exclusively.
• 6:24
  Our source actually uses Grokbot for emails and
• 6:27
  loads it, but they joined the Duet wait list
• 6:29
  anyway to test them head to head. The curiosity
• 6:31
  is too strong. So what does this all mean? It
• 6:34
  means you have to budget your AI interactions.
• 6:37
  Do I get charged tokens even when I'm not actively
• 6:39
  typing to it? Yes. Every time it wakes up to
• 6:42
  check your apps in the background, it eats into
• 6:44
  your shared subscription limits. Ah. So background
• 6:47
  convenience could quietly drain your daily allowance.
• 6:50
  Exactly. You might wake up and find your limits
• 6:52
  are already capped. We're going to take a quick
• 6:54
  pause right here. Mid -roll sponsor read inserted
• 6:56
  here. All right. Let's pull all this together.
• 7:00
  We're witnessing a major shift. We really are.
• 7:03
  We are moving from AI as a reactive tool you
• 7:06
  talk to to a proactive employee that works while
• 7:09
  you sleep. Claw Duet is a massive leap in workflow
• 7:13
  integration. But it requires careful management.
• 7:16
  You have to monitor your security and your token
• 7:19
  usage. It's a supervised partnership, for sure.
• 7:23
  I want to leave everyone with a thought here.
• 7:25
  If an AI is constantly reading your emails, managing
• 7:28
  your payroll, and catching your scheduling mistakes
• 7:30
  in the background. At what point does your digital
• 7:34
  agent understand the reality of your business
• 7:36
  better than you do? It's a wild thought. It sees
• 7:43
  the patterns you're too busy to notice. Which
• 7:46
  AI agent would you trust to run your daily workflow?
• 7:49
  Keep exploring. Until next time, Outro Music.