Good morning. Here's what's worth acting on today.

BREAKING STORY

From December, ChatGPT Plus, Pro and Business subscribers will draw from one shared allowance across ChatGPT and Codex, according to reports by VGTimes and SQ Magazine citing OpenAI's help article, Understanding ChatGPT and Codex usage limits. The meter covers Codex, OpenAI's coding agent, and Work, the agent mode where ChatGPT carries out long tasks on its own. Quick answers in Instant mode don't count. Remaining usage and reset times will appear under Settings, then Usage, and if you run out you can wait for the reset, upgrade or buy credits. OpenAI hasn't published the size of each plan's allowance or an exact start date.

If one subscription covers your coding and your admin, a heavy Codex day now eats into the budget for your research, writing and paperwork. Startup Fortune calls it "a real change to what a subscription actually buys", and the risk for a solo business is running dry mid-week with client work still open.

My take: Before December, log for a week which jobs you run in Codex and which in ChatGPT agent mode, and move the repetitive ones to the API or a cheaper model so one heavy coding day doesn't lock you out of everything else.

MARKET TRENDS

Upwork's third-quarter results show a split market. Work bought from freelancers with AI skills reached an estimated annualised run rate of nearly $450 million, up 55% year on year, it says. Total gross services volume (the value of all work bought on the platform) fell 10.1% to $914.1 million. Upwork also says AI agents have made nearly 50 million requests through its MCP server, a connector that lets AI assistants search and hire on the platform, with weekly active users up 20 times since its August launch.

How to make money from it: The money is following AI skills, and some of the buyers are now software. Turn one service into a fixed-scope AI package, such as setting up an agent or automating a weekly report, with a clear price, scope and turnaround. Put the AI skill in your profile title, and write listings plainly enough that an agent searching for a client can match them.

People assume that real businesses get built by other people with more resources. It's how great ideas get stuck in notes apps and conversations.

But you don't need six months and a developer's salary to validate your concept. You just need a weekend. Describe what you want to exist and Lovable builds it: a real site, a real product, ready to accept real payments from real customers. Not a mockup to show people. An actual thing they can use.

You don't need anyone's sign-off to get started. Take $5 off your first month.

Find out this weekend

_{^{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.}}

GROWTH HACK / DIY

The idea: Your heaviest solo users are the likeliest to bring in a team, and what stops them is rework: rebuilding projects and re-explaining context. The fix is a one-click move from a personal space into a team workspace that keeps each item's history.

v0's Teams documentation shows the pattern: you can move one chat or many into a team, a linked Vercel project goes with it if you own that project's current team, and team owners see the full history, including messages from before the move. Each seat gets $30 of monthly credits that roll over for one month, with a Shared Credit Pool used once those run out. The v0 changelog says personal accounts became team workspaces on 6 October 2026.

Do it today:

What you'll need: a "change owner" action, a bulk-move screen and a confirmation that says what moves and who will see it. Budget roughly one to two weeks of build for a small app (our estimate). Copying instead of moving carries a cost: Notion's guide warns that duplicated pages can lose links, permissions and page history.

Full playbook →

AGENT LAB

Most agents get a master key at the start and keep it forever. Making the agent ask before every action is safer, but you soon click "yes" without reading. Vercel describes a middle path in How Vercel Agent accesses production safely (23 Oct 2026).

How it works: The agent investigates with read-only access. If it needs more, it writes a plan listing each task, why it is needed and the named permissions (scopes) it requires. Unknown or malformed scopes are rejected before you ever see the plan. You approve once, and Vercel issues a shorter-lived token covering only the read baseline plus the approved scopes. Before each use, the system checks the plan is still active, so pressing Stop cuts access even if the token has time left. Approval can never grant more than you already have.

Try it: Give your agent a read-only key by default. Make it output a JSON plan of tasks, reasons and scopes, and check every scope against a short allowlist in code. On approval, issue a narrow key: a GitHub App token that expires after an hour, or a Stripe restricted key held back until a matching plan is active. Then test that Stop really stops the next call.

Full breakdown →

TRENDING AI RESEARCH

What it says: When an AI assistant acts without being asked, its most common mistake is being over-eager. In a new study called HALO, 20 experts graded 8,000 unprompted responses from 10 models across 400 everyday scenarios. Of the responses they rejected, 61.20% came too early or did too much. The best proprietary model was acceptable 72.88% of the time; the best open-source model, 56.63%.

Why it matters: If your product sends reminders, nudges or "I noticed…" messages, or lets an agent act on someone's inbox or calendar, every unwanted interruption spends trust and every unwanted action can cost money. Even the best model gave an unacceptable response about one time in four in this test.

How you can use it: Give every unprompted feature an escalation ladder: mention, then suggest, then draft, then act. Start on the lowest rung that helps. Only speak up when the agent can point to a specific record showing a real need, aim just before the need is due, and let the system act on its own only where the person has switched that on and can undo it. Using an assistant yourself? Set it to ask before it acts.

Full breakdown →

TOOL REVIEW

What it does: Raah.dev, which launched on Product Hunt on 1 November, puts your analytics, Stripe revenue, Core Web Vitals, API latency and errors behind one cookie-free script. It matches each payment to the visit that produced it, so you can see which channel or landing page actually earns money. It also offers email and webhook alerts, public status pages, and an MCP server so coding agents can query your data (it says availability is limited).

Price: As of 3 Nov 2026: free for one site, Pro at $9 a month for five sites, Premium at $25 for ten. Native Slack alerts are still listed as upcoming.

Who it's for: Solo founders and small teams running a few SaaS or indie sites without an ops person.

Verdict: Try it if you run a few small SaaS or indie sites and want revenue, traffic and speed in one cheap view. Skip it if you already pay for PostHog or Plausible plus Sentry.

Full review →

What Got Funded Yesterday

QUICK HITS

That's today. Hit reply and tell me what you're building. I read every email.

Dev