Good morning.

In the last four weeks, Notion, Meta and Zapier each gave their agents more room to work without you, and Anthropic shipped a Claude that costs less to run and can drive a browser.

Almost every one of them shipped a second feature right beside it:

A point where the agent stops and waits for your yes.

In today’s The Scane issue: eleven releases, each with my read on what it means for a business your size, and a straight call on each one:

Use it now, wait, or ignore.

— Sam

Let’s get into it.

Notion rebuilt its agents on October 19. They now get their own computer and browser, and a single task can run for up to 24 hours, where the limit used to be five minutes. A Review for me check reads the agent’s work before it’s done.

Notion also moved its AI to credits: the Notion Agent comes with the Business plan at $20 per member a month, each seat includes 1,500 Standard and 500 Select credits a month, extra Custom Agent credits cost $10 per 1,000, and existing customers move to the new credits on November 2. (Source)

A day is long enough for an overnight job: read the week’s client notes, pull out the open questions, and leave a brief on each client’s page by morning. What decides whether it’s worth it is the credit cost of one run.

Pick one job whose result you can check in five minutes, run it once the new credits apply, and read what it cost before you give it the backlog. Free and Plus get a limited trial, enough to test that one job.

Verdict: Wait. Existing plans move to the new credits on November 2, so price one run before you hand it real work.

Anthropic released Claude Opus 5.5 on September 22. On the API it costs $4 per million input tokens and $20 per million output tokens, and Anthropic says typical workloads cost about 40% less than on Opus 5, a figure it measured itself. It reads files, PDFs and images, and through the API it can use a browser and a computer on its own. For the Claude apps, the launch raised the usage limits on the Pro and Max plans. (Source)

In Claude’s desktop app or Cowork you don’t pay by the token, so the higher limits mean longer sessions before you hit a wall. The browser and computer control matter most where a tool has no connector, like the supplier portal or the old booking system you log into every week. Start with one job you can undo, like pulling last month’s invoices into a sheet, and check every row before you trust it with anything that sends or pays.

Verdict: Use it now. It’s in the plan you already pay for, with more room to work.

On October 6, Zapier folded its standalone Agents into an AI by Zapier step inside the Zap editor. The step can reason, call tools and finish a task on its own, and it sits beside the triggers, filters, Paths and fixed actions you already use, with every run in Zap History. You can require approval for individual tool calls.

It bills in tasks: a stronger model uses more tasks per run, tests use tasks too, and a run pauses at 75 tasks unless you raise the cap. Existing Agents have a migration path, and Zapier hasn’t set a date to retire them. (Source)

A lot of you said connecting your tools is where you’re stuck, and this is the setup to copy:

Let the model handle the one messy step, and let fixed rules handle everything after it.

For a coach, a new inquiry arrives, the AI step reads it against your offers page and labels it, a Path sends good fits to a reply draft, and the approval holds the reply until you’ve read it. Watch the task count on the first ten runs before you switch it on for every inquiry.

Verdict: Use it now. If you already pay for Zapier, it’s the safest way to put an agent into a live workflow.

Meta launched Muse for Small Business on September 29, in the US and Canada. It connects your Facebook and Instagram business data to tools including Shopify, Stripe, Klaviyo, QuickBooks, Notion and Canva. It can plan growth, analyze your campaigns and inbox, draft campaigns, review finances for unusual expenses, and draft replies to the emails it flags. Meta says nothing publishes, sends or spends without your approval. It’s free for most uses, with paid plans Meta hasn’t priced publicly. (Source)

For a store or a coach running Meta ads, this is the first assistant that sees the ad account and the sales side by side without you building the connection. It’s also Meta’s assistant reading Meta’s ads, so check its numbers against Stripe or Shopify before you act on a recommendation. Connect it read-only, ask for one Monday brief covering sales, ad spend and anything unusual, and compare it with your own numbers for two weeks.

Verdict: Use it now, if you’re in the US or Canada and run Meta ads. It’s free, and the approval rule is built in.

OpenAI started rolling out ChatGPT Slides on October 16 to Pro, Business and Enterprise users with access to Spaces. You start it in ChatGPT Work, inside a Space, from a conversation, files, an outline or an example deck. You can edit together, leave comments and ask ChatGPT to apply the changes, and on the web, decks export to PowerPoint, PDF or Google Slides. It isn’t on Plus or Free. (Source)

For a consultant, the decks that make money are the proposal, the workshop and the quarterly review, and each starts from call notes, numbers or an old deck you liked. Give it your best past deck as the example, so the layout and voice are yours, then check every number and price before it goes out. A client can mark what’s wrong in the comments, and ChatGPT applies only the changes you agree to.

Verdict: Use it now, if your plan has it. On Plus, wait for the rollout.

Shopify says its Agentic Storefronts are active by default for eligible stores, which puts products in front of buyers inside ChatGPT, Google and Meta’s Muse. To sell through ChatGPT, a store has to sell to US customers (it can be based anywhere), have products that qualify for Shopify Catalog, accept Shopify’s supplemental terms, and have its terms, privacy and refund policies filled in (Shopify’s requirements).

In an October 6 case, Shopify said Cottonique’s AI-referred sales rose 276% year over year and its total sales 52%, with no base, period or method published. (Source)

In August’s Scan, AI-referred buyers converted better and brought in a sliver of revenue (Your Best Buyers Are 0.1% Of Revenue). The new part is that your store may already be listed. An assistant can only say what your product data says, so before Black Friday, open your top ten products and check the sizes, what each one is for, who it isn’t for, shipping and returns. Then make sure your reports separate AI-referred orders from direct ones.

Verdict: Use it now. It’s already switched on, so the only choice left is whether the data behind it is right.

TinyFish, an API company, reports reaching 300,000 users this month (users, not paying customers): the first 100,000 took four months, the next 100,000 eighteen days, and the last 100,000 ten.

Its marketing team published the content workflow it runs every weekday. It checks for new blog posts, pulls out the text, images and links, and asks Claude for separate Medium and LinkedIn drafts. Then it checks every number and link against the original post, flags anything the post doesn’t support, and sends the drafts to Slack, where a person edits and publishes them. The team calls the milestone a company-wide effort and doesn’t credit it to this workflow. (Source)

The part to copy is the check, because a draft that quotes your post can still invent a statistic. Put that check before the human step in any repurposing setup.

You can build it in Zapier with Claude: a new post goes up, Claude writes the LinkedIn version and the email version, a second step lists any claim that isn’t in the original, and both drafts go to your inbox. One post a week becomes three channels for the cost of a read.

Verdict: Use it now. It’s an afternoon to build, and the claim check is the piece to build first.

On September 11, Salesforce made four Agentforce agents generally available, for support, shopper discovery, inbound pipeline and customer service, and said its outbound agent, Hunter, becomes generally available in November.

It published customer figures: 50% of Engine’s chats fully resolved, 90% of Hibbett’s core shopper journeys handled, four times the conversation volume driven by Asana’s website agent, 60% of Perk’s pipeline built by Hunter, and 79% of Anthropic’s conversations resolved by Fin, the former Intercom, which Salesforce agreed to buy in June. None comes with a definition, a baseline or a denominator. (Source)

It’s enterprise Salesforce with no published pricing, and the figures can’t be compared with anything you’d measure. Read it as a list of the jobs worth handing over first: answering the same support questions, qualifying inbound leads, and chasing a deal that went cold, all of which run on tools you already pay for.

Verdict: Ignore. It’s built for companies far bigger than yours, and the numbers don’t transfer.

HeyGen launched Video 1.0 (heygen-video-1) on September 30. From a text prompt, a first-frame image or reference media, it returns a 5-to-15-second clip with generated dialogue, ambient sound and effects. It’s only available through paid API keys or routes like OpenRouter, and it isn’t part of HeyGen’s app plans.

Launch pricing is $0.01 a second through October 31. HeyGen’s own docs give two different regular prices at 768p ($0.02 and $0.03 a second), and 1080p and 2K cost three times as much. (Source)

For a store, a ten-second clip of the product on a bathroom counter, with the sound of the jar opening, is a fair ad test, and at launch pricing a batch of twenty costs a few dollars. The render is the cheap part: checking the hands, the label text and the product against the real thing, frame by frame, is where the time goes.

If you’d rather work inside an app, HeyGen’s avatar Video Agent has been on its Creator and Pro plans since January, at $29 and $49 a month.

Verdict: Wait. It’s API-only for now, and the regular price isn’t settled.

In an October 19 post, Shann Holmberg described SageCut, a quality-check skill built for videos made in code with Remotion. Cheap automatic checks and a cheaper vision model look at the frames first, and only the defects they flag go to Opus, which investigates and changes the code.

Holmberg reports 90% lower review cost and 75% less review time, and the builder reports QA cost per video falling from $127 to $0.16. These are self-reported figures, and it isn’t clear whether the skill is public. (Source)

Unless you render video in code, skip the tool and keep the pattern: send the expensive model only the failures. If Claude reviews your weekly report or a batch of posts, put a cheap first pass in front (word counts, links that open, numbers that match the source) and hand Claude only what failed.

Verdict: Ignore. The tool is for code-rendered video, and the pattern is the part worth keeping.

Sunra launched an MCP connector that brings its catalog of image, video, audio and 3D models into the assistant you already use: Claude and ChatGPT connect with a sign-in, and Claude Code, Codex and Antigravity with a setup prompt.

Jobs draw on the same credits as Sunra’s web app, show an estimated charge before they run, and wait for your confirmation unless you turn on auto-confirm. Sunra says that where a model’s maker publishes an API price, it charges that price. Plans start at $12 a month for 1,200 credits. (Source)

If you already pay for two or three image or video tools, this puts them behind one balance and one chat, which suits an agency making ad variations or a creator making thumbnails every week. Keep the confirmation on, because an agent asked for “a few options” can ask for a lot more than a few. Run one small quoted batch, compare the cost of each image you’d keep with what you pay now, and decide from that.

Verdict: Wait. Try it only if you already pay for more than one generation tool.

One tap tells me which of these to go deeper on next. If it’s something else entirely, pick None yet and hit reply.

Five of these releases let an agent do more without you, and four of the five shipped a checkpoint beside it:

Notion’s review, Zapier’s approval on each tool call, Muse’s rule that nothing publishes or spends without you, and Sunra’s price before each job.

The vendors are telling you where these things still go wrong:

The moment an agent acts outside your own files, by sending, spending or publishing. That’s where your one approval belongs, and the rest can run.

In Cortex, I share the formulas and strategies that help busy online business owners implement AI and Agents inside their businesses, in just minutes or hours instead of days or weeks.

Learn more about Cortex here

Talk soon,
Sam Woods
The Editor