# The Edge: Let your agent pay? Plus Claude Code's new coordinator

> For ChatGPT Pro users in the US, whose dot (one of OpenAI's agents in ChatGPT) can now check out, and for Claude Code users running parallel work.

- Published: 2026-11-16
- Publisher: Rig Labs (https://riglabs.ai/)
- Canonical: https://riglabs.ai/newsletter/2026-11-16/

---

Issue 6. Our read: what stopped people from letting an agent shop was the moment of typing the card in. OpenAI now has you approve every purchase with Face ID or Touch ID and keeps the card number away from the shop, which removes that moment. The harder question is how long people keep reading what they approve.

In this issue:

1. ChatGPT Wallet: what a dot (one of OpenAI's agents in ChatGPT) can now buy for US Pro users, what it adds over checking out yourself, and our two bets
2. Guide: set up Wallet and make a first small purchase without surprises
3. Claude Code Projects: a project splits your coding tasks across threads, each a separate Claude session, and its coordinator now reports on every thread. Plus the settings to check before a project uses up your plan
4. What's new: themes for the Claude Code desktop app, Anthropic's checklist for always-on agents, and a browser-use update that costs more by default

#### 01 / The big one

## What changes when the agent holds the card?

OpenAI launched [ChatGPT Wallet](https://openai.com/index/introducing-chatgpt-wallet/) on November 13, for ChatGPT Pro users in the US, and it is rolling out gradually. You save one credit card, your dot (one of OpenAI's agents in ChatGPT) finds what you asked for, and you approve each purchase with Face ID or Touch ID. The payment network then issues a stand-in card number for that purchase, so neither the dot nor the shop sees your real card details, and you keep your card's rewards and its protections for eligible disputes.

> **Guide:** [Let your ChatGPT dot pay for what you ask it to buy](https://riglabs.ai/guides/2026-11-16-let-your-chatgpt-dot-pay-for-what-you-ask-it-to/)
> For: ChatGPT Pro users in the US, 18 or older, who use a dot. Needs: a US-issued American Express, Chase, Mastercard or Visa credit card (no debit or virtual cards).
> A step-by-step tutorial: save one card, let your dot shop, and approve each purchase with Face ID or Touch ID. The setup file lets your AI assistant walk you through it.

Until now the usual flow was: the agent finds the hotel or the charger, then you open the shop's site, type your card and check out yourself. Wallet moves that last step into one approval card, which is where you approve the purchase with Face ID or Touch ID. It also has limits a browser checkout does not: you must be 18 or older with a US billing address and a US-issued American Express, Chase, Mastercard or Visa credit card (Chase only for the primary cardholder, with approval in the Chase app), Wallet works only in dot conversations, debit cards and virtual cards are out at launch, Wallet may not complete checkout at every store, and "ready to check out" does not mean the shop placed the order ([help page](https://help.openai.com/en/articles/20001654-making-purchases-with-chatgpt-wallet)). So who gains when checkout becomes one approval? The bold lines are our forecasts.

- **Shops that block agents will lose small, repeat orders to shops that let them through.** The reason is effort: once an agent can check out with one approval, a store that sends the human back to its own site is the slower option. OpenAI itself says websites need to make sure agents are not blocked by mistake. Of course big retailers have reasons to keep the customer on their own pages. If by spring no large retailer that blocked agents has opened its checkout to them, we are wrong.
- **People will stop reading the approval card within weeks, before OpenAI ships pre-approval, its planned way to let users approve purchases in advance inside limits such as a spending cap or a category.** The approval card works only while people read it, and repetition kills reading. If OpenAI ships pre-approval and people still read the card, we are wrong. Watch for the first story of a subscription someone approved without reading its terms.

Paste this into ChatGPT, in a conversation with your dot:

```
Show me my saved Wallet cards. If Wallet is not available on my account yet, say so.
```

Your move: make your first purchase something small you already planned to buy, and read every line of the approval card. [Who can use it and what can go wrong](https://riglabs.ai/blog/chatgpt-wallet/).

[Give it to your agent](https://riglabs.ai/guides/2026-11-16-let-your-chatgpt-dot-pay-for-what-you-ask-it-to/#agent)

#### 02 / Coding

## Claude Code Projects: one coordinator for parallel threads

For: Claude Code users on Pro, Max or Ultra who code in GitHub repositories and run several tasks at once. Needs: Projects in your sidebar (public beta, rolling out gradually).

On November 13, Anthropic made the coordinator in [Claude Code Projects](https://code.claude.com/docs/en/claude-projects.md) more proactive, per [@ClaudeDevs](https://x.com/ClaudeDevs/status/2121375309287301360): it now reports on every thread of a project. A project is one Claude conversation that splits your coding tasks across threads, each a separate Claude session on its own Git branch. The coordinator now answers your questions in the project chat and posts updates about its threads there.

How is it different from opening several Claude Code sessions yourself? You ask one place where things stand instead of checking each session. The cost is real, though: a project can use up your plan. Per Anthropic's docs, every thread in a new project runs on Opus at high effort until you pick another thread model, and a thread that stops at your usage limit continues on its own when the limit resets, unless you change the auto-continue setting. Team and Enterprise plans do not have Projects yet, and a project belongs to one user, so you cannot share it.

Paste this into the project chat in Claude Code:

```
Give me a status update on every thread. From now on, only post when something finishes or is blocked.
```

Your move: before your next project starts work, open Project settings > General, and pick the thread model and the auto-continue setting you want. [The settings and who can use it](https://riglabs.ai/blog/claude-code-projects-coordinator/).

[Give it to your agent](https://riglabs.ai/blog/claude-code-projects-coordinator/#agent)

#### What's new

### Claude Code desktop themes

For: people who use the Code tab of the Claude desktop app. Needs: Claude Code v2.1.324 or later for colors, local sessions only.

Since November 13, a mod can change the colors of the [Claude desktop app's Code tab](https://code.claude.com/docs/en/plugins/mods/interface.md) and draw an animated background behind the chat. You describe the look and Claude writes the mod.

For you: one sentence gets you a theme. Read the files Claude writes before you approve them. [The prompt to paste](https://riglabs.ai/blog/claude-code-desktop-themes/).

[Give it to your agent](https://riglabs.ai/blog/claude-code-desktop-themes/#agent)

### Anthropic's guide to always-on agents

For: anyone who runs an agent on a schedule or in the background. Needs: nothing.

Anthropic published [Building always-on agents](https://claude.com/resources/articles/building-always-on-agents) on November 13. It says an agent that seems always on sleeps most of the time, and needs five parts around it: an event inbox, timers, memory, permissions and an outbox.

For you: a checklist to find the part of your own scheduled agent most likely to break. [The five parts and a review prompt](https://riglabs.ai/blog/always-on-agents-anthropic-guide/).

[Give it to your agent](https://riglabs.ai/blog/always-on-agents-anthropic-guide/#agent)

### browser-use 1.0.6

For: developers who use the browser-use Python library with the Browser Use gateway and do not name a model. Needs: a one-line edit to keep the old price.

[browser-use 1.0.6](https://github.com/browser-use/browser-use/releases/tag/1.0.6), released November 16, runs BU 4.0 by default when your code names no model. Per the release notes, BU 4.0's tokens cost 15 times as much as those of bu-fast, the old default.

For you: if your code names no model, add `llm='bu-fast'` before you upgrade to keep today's cost. [The one-line fix](https://riglabs.ai/blog/browser-use-bu-4-default/).

[Give it to your agent](https://riglabs.ai/blog/browser-use-bu-4-default/#agent)

**Would you let an agent buy something for you this week, and what would it be?** Tell us what you would hand over first, and what you never would. Reply to this email.
