Claude Duet is out, and explaining how it fits required something NLW hasn't done in a while: a full tour of his actual AI stack. Strategy in Opus 5.5, research cross-checked between Astra and Opus, a Claude-built production pipeline behind the website — and a first honest verdict on whether personal agents are chiefs of staff or just proactive baby-yous.
Personal agents are still an experiment — chief of staff, or proactive proxy employee?
The form factor has traction — Muse is growing, and we've been playing with this shape since OpenClaw exploded — but the open question is what these things are actually for. Are they the master coordinator of everything you do with AI, or a very specific proactive proxy-you that figures out how to be helpful without being asked? A few days into Claude Duet, NLW's answer: much less a shift in how he interacts with AI, much more an easy-to-set-up little baby version of himself that does the stuff he doesn't want to do. Valuable — but future-of-the-industry valuable remains to be seen.
Yes, there's early traction — especially with Muse — and the form factor has been around since OpenClaw exploded earlier this year. But the unresolved question: are these agents a chief of staff organizing everything you do, or a very specific proactive proxy-you that goes out and figures out how to be helpful without being asked?
NLW's main job across AIDB and Superintelligent is deciding what to do or create next, and he thinks by talking it out — reams of 'Walk Notes' in Notion from walking, driving, and thinking sessions. Across every AI session he's ever had, roughly 60% of time spent lands in this strategy realm, even if building consumes more tokens.
For a long time, models weren't much better than writing on paper because every idea was the best idea ever. For strategy work you don't want a yes-person — you want critical engagement grounded in actual reasoning, not a model trained that pushback is good. Dynamic, creative, adaptable, and confident is a really hard line to walk.
It displaced Fable 5.1, which was okay, and before that OpenAI models — though for most of this year the GPT models have felt flat and insufficiently dynamic for this sort of work. Every major new model gets tested against strategic externalization to see if it's a one-to-one replacement.
Because research ends up said publicly on AIDB (and on the Love Murder true crime podcast), sourcing and hallucination-resistance are non-negotiable. NLW uses Astra and Opus about equally — and on important topics runs both, then has each review the other's output as an extra confidence layer before it becomes content.
The nightly process: NLW surveys the day's news and discourse, a researcher in Australia — a hack that makes the operation effectively 24-hour — builds an overnight dossier of sources and key social commentary, and NLW lays out screens and lets it rip, unscripted. The 'AI talking about AI' comments are funny because there's remarkably little AI in the pipeline.
Episode descriptions and social posts run through AI daily, and titles are a 50/50 back-and-forth where the AI gets a chance to argue him off his draft. Even if Opus or Fable were as good or better, the long accumulated track record of likes and dislikes makes switching not worth the tuning time — plus OpenAI's models get credit for concision.
Daily cover images and in-show visuals have mostly consolidated on the GPT image models — not necessarily because they're better, but because the headline-to-cover workflow lives in one thread. Ideogram's four-at-a-time generation with variance toggles wins for iteration; Midjourney's strange taste wins when you just want something visually arresting and can give the model latitude.
From podcast statistics to the tax-extension ritual of dumping in a year of financial data and having it extract expenses, almost every analysis project runs through Astra in Codex — the best results plus a concision of thought that suits the work.
After cycling through OpenClaw, Codex, Cloud Code, and Cursor, the organizing, narrative architecture, and orchestration all happen in Claude Code — with Codex-plus-Astra reserved for builds that need native access to the GPT image models, like visual companion experiences. Claude Code also gets GPT image access via API, though results differ from using the models natively inside Codex.
It had been easy to start building things for a while — the activation energy was low — but projects hung at 80 or 90% complete. Fable brought the completion energy down to match, making the difference between starting things and actually getting them done.
The mandate: an information-dense show creates a high barrier to entry, and listeners wanted an easy way to share specific parts with peers. So the site ingests each episode's transcript and extracts it into shareable web assets, with a no-image-generation aesthetic built in Fable and Claude Design. The entire first draft was completed the afternoon NLW got Fable access — the day before it was taken offline in the export-control dust-up with the White House.
A secondary pipeline turns each episode's extracted cards into tweets and a LinkedIn post, scheduled through Typefully. NLW explored wiring directly into the platform APIs, but a third party had already done that work — plugging in was way easier.
Especially with Opus 5.5, the models' video-editing jump is big enough to be worth exploring on its own. NLW's pipeline feeds Claude Code each day's transcript; it judges the most compelling self-contained 60-to-90 seconds, extracts the clip, builds visual assets around it, and packages it for upload.
A 45-minute-to-hour weekly show tracing the social, cultural, technological, and economic decisions behind today's biggest stories — born from AIDB never having space to go deep enough. First three episodes: how America fell out of love with tech, how stuff got cheap while the important things got expensive, and how we started betting on everything via prediction markets.
After recording, the raw transcript plus original research goes back to Opus 5.5 for a first-draft companion essay — the same argument as the show, built for readers, with links and sources. The essay then goes into Codex with Atlas and GPT image for a full look and feel, followed by an exhausting edit in which NLW rewrites about half of it.
The OpenClaw era experiments — individual claws on heartbeats and cron jobs, a chief-of-staff agent coordinating them, a persistent research agent monitoring AI adoption studies — all got shut down because they didn't add enough value to justify maintenance. Part of that is personal: 90% of NLW's calls are unscheduled and his inbox system is starring and clearing every couple of days. The new question is whether this generation lowers setup cost enough to flip the math.
The canonical consumer pitches — flight booking, agentic shopping, restaurants — don't fit someone who books within one airline's miles ecosystem: explaining the variables and contingencies to an agent would take exactly as long as using the airline's own interface. The professional side of personal agents was always the only live question here.
Instead of chats scattered across projects that each hold only their own context, one thread holds everything, pulls up what it needs, and spins up sub-agents on demand — removing the cognitive load of organizing your own AI usage. This only became viable with back-end advances in compaction and context-window management; early-2026 training programs literally taught handoff documents as a workaround.
Third-party agents like Instinct and Town lose on the unwillingness to move work context out of ChatGPT and Claude; Muse skews too personal for now. Grokbot earns a 'stay just engaged enough to not start from zero' allocation for its X integration. Google isn't worth heavy investment yet — though NLW is optimistic that changes — and Microsoft is tracked professionally but skipped by a solo operator biased toward nimble tools.
After a full day of walking-and-hiking strategy with Opus 5.5, NLW shared the entire six-hour conversation with his dot, Quasar. Its response: 'Great, this information has been saved for the future.' As a feature it adds little to ChatGPT's reactive strengths and is very weak on proactive work — though it could have value as a master holder of context you dump once instead of repeating across chats.
Anthropic's work-focused, personality-light agent asked to connect Slack, Gmail, and Drive immediately and proposed a stack of proactive projects before anything else. One of them — policing podcast ad-slot double-booking — was a false alarm it reminded NLW about four times before he killed the entire workstream. Being able to be proactive about something isn't the same as it mattering.
For exactly the kind of email NLW would star and return to days later, Duet has proposed responses he can edit or authorize straight from chat — a couple sent as-is, and the most substantive creative proposal got him 90% of the way without going back to think about it. Plus a surprisingly good web interface with channels, artifact integration, and automatic task grouping. A few days in, this is the most optimistic he's been about the form factor.
The moment a feature works in one of these products, it becomes part of all of them within weeks if not days. If Dots' lack of proactivity is a common complaint, OpenAI will tweak it; if people prefer Duet's UI, the other labs will copy it.
Personal agents haven't become the main way NLW interacts with AI. They're an easy-to-set-up little baby version of himself doing a bunch of stuff he doesn't want to do — which is valuable, but whether it's future-of-the-industry valuable remains to be seen.