Anthropic bets on non-code agents & OpenAI wants your wallet | Intent, 0027

Oh wow, it’s been a minute!

We haven't sent an issue of Intent in a while -- hope you've been swell in the meantime!

As a reminder: Intent exists to help tech talent become more informed, more fluent, and more aware of the forces shaping their careers. Hit reply with feedback or hot takes.

The agenda ahead

1. Anthropic's Sonnet 4.5 is nice, but Claude Code 2.0 is a bigger deal

2. OpenAI's new Sora app is neat, but "Buy it in ChatGPT" is a bigger deal

3. Quick-hits include: AI jewelry, California AI policy, and AI inserts itself in your work

Programming note — the next cohort of our 4.9/5.0 star cohort-based course, AI Fluency Bootcamp, begins on Oct. 20! Early bird special for the next 48 hours: use code AGENTIC for 50% off. Designed for operators, marketers, designers, PMs, GTM, and engineering leaders who want compounding AI leverage at work.

Anthropic releases Sonnet 4.5 and Claude Code 2.0

This week, Anthropic released Claude Sonnet 4.5 and a meaningful upgrade to Claude Code (plus, Claude Agent SDK). Sonnet 4.5 is now their best model. It scores well across benchmarks for software engineering, domain knowledge, and alignment, but in early feedback and testing, power users think it's about as smart as GPT-5 Thinking for general chat and not as good as GPT-5-Codex for code.

But the Claude Code updates are paying attention to! It's a set of upgrades to their agentic tool that runs in your computer's terminal, meant primarily for AI code generation. But because it lives on your computer, can see your files, execute commands, edit local data, and interact w/ your operating system, AI enthusiasts have been using it as a general purpose copilot.

And Anthropic is now acknowledging & embracing these non-code use cases directly:

The pattern to watch: “general‑purpose copilots” are going to be the breakout, not agents that live within each particular application that you use. That's what makes Claude Code so effective at being general-purpose: it's cross-application, can run terminal commands to achieve lots of different ends, and is willing to follow non-code instructions.

The agentic container means lots of autonomous task capability: editing CSVs, wrangling docs, doing deep web research, calling third party APIs, converting media.

A real-world example: Let’s say you need to download and transcribe a YouTube video. Manually, this takes 10 minutes, a few different tools, and some technical know-how.

With Claude Code, you simply ask: "Hey, download this YouTube video [url] using yt-dlp, extract the audio using ffmpeg, and then call the ElevenLabs API to build a diarized transcript." Go make a cup of coffee and by the time you’re back, done!

(and if you didn't already know about yt-dlp, ffmpeg, and ElevenLabs, you'd simply start by asking Claude what tools and libraries it would use to achieve your objective -- the flexibility is the point!)

While tools like Claude Code (and OpenAI’s equivalent, Codex CLI) require a bit of persistence to set up if you aren't technical, the juice is worth the squeeze. They are a preview of the agentic frameworks that will infect every workflow and functional area in the next 6 to 12 months. The era of the general-purpose copilot has begun.

Want a walkthrough? Join Sherveen’s free live session: How to use Claude Code or Codex CLI for non‑code use cases. RSVP here.

OpenAI rolls out "Buy it in ChatGPT"

It's long been rumored that OpenAI wants to experiment aggressively with business models beyond the ChatGPT subscription. They were recently spotted posting a job for an ad platform engineer, and now they're enabling ecommerce directly in the app.

Instant Checkout in ChatGPT allows users to purchase products without leaving the conversation, leveraging early partners like Shopify, Etsy, and Stripe to power the transaction for integrated merchants. Think: an agent that reasons over a cart state, calls merchant tools, and completes a purchase without kicking you out to a dozen tabs.

They also rolled out related infrastructure in the form of ACP, the Agentic Commerce Protocol, an open spec for AI agents to transact with businesses.

For years, LLMs recommended products -- but then punted you to the open web. This is one of the first mainstream, integrated agent‑to‑merchant flows from a frontier‑model vendor.

The critical tension here is trust. Users rely on LLMs for (relatively) unbiased information and recommendations. If ChatGPT’s advice becomes skewed by affiliate fees or sponsored placements, that trust might go poof.

More broadly, this is the AI industry’s first major foray into integrated agentic payments. Just as agents are getting good, frontier labs will want to let them "do business," to put it simply.

But one of our hot takes… there’s a real chance this could be better than our current e-commerce reality. Instead of visiting Amazon and navigating a minefield of sponsored links, questionable manufacturers, and fake reviews, an AI agent you trust could synthesize genuine reviews, find the best product for your specific needs, and execute the purchase.

We'll all need to keep our eyes peeled to make sure incentives don't corrupt results!

Small bites

• AI jewelry that's... actually jewelry. Two female founders launched Taya, an AI pendant designed as jewelry first (aka: it looks good!) with on-board intelligence to keep your memories, reflections, and ideas. A refreshing counter-aesthetic to the typical "ugly square on a lanyard" we've seen so far.

• Meanwhile... AI pendant Friend got the attention it asked for: Controversial startup Friend (see an early review here) spent >$1M on NYC subway ads for its AI pendant, and... NYC doesn't like it.

• OpenAI announced GDPval -- a new AI eval suite focused on 'economically valuable tasks' across 44 occupations rather than abstract benchmarks.

• SB 53 is law in California -- the Transparency in Frontier AI Act requires frontier AI labs to disclose safety testing and disclosures + adds incident reporting and whistleblower protections.

If you only do one thing this week

Install an agentic CLI (Claude Code or Codex CLI) and give it a real, multi‑step job with a clear definition of “done.” You’ll feel the future snap into place. If you run into roadblocks with the set up, just ask your favorite LLM to give you step-by-step help!

(oh, and give Sora 2 a try)

Think a friend could use a dose of Intent? Forward this along – good emails can turn a week around.

Sent with Intent,
Free Agency