PROMPTWIRE
Gartner went through the thousands of products being sold as AI agents and found that only around 130 vendors were offering anything genuinely agentic (Gartner, via Futuramo).
Everything else is a chatbot with a better landing page.
The difference is not academic, and it is not about how clever the thing is. A chatbot answers. An agent finishes. It writes back into a system, completes a job without you prompting each step, and hands off when it hits something it should not decide. Mixing those two up costs money in both directions: people pay agent prices for a search box, or they write off agents entirely because the demo they saw could not actually do anything.
Ten minutes below tells you which one you are paying for. Then the part that matters: building a real one out of tools you already own, instead of buying another subscription.
⚡ This Week In One Minute
Workflow: Test whether the AI tool you pay for is actually an agent. Then build a real one from what you already have.
Big Story: 17% of organisations have deployed agents. More than 60% plan to within two years. That gap is the steepest adoption curve Gartner has ever measured.
Tool of the Week: ChatGPT Dots. Always-on agents with their own cloud computer, launched at DevDay last week.
In Pro: the build, end to end. The job to pick, the context that makes it work, the write-back permission, the escalation rule, and the six checks before you let it act.
🔁 This Week's Workflow
Find Out If You Are Paying for an Agent or a Search Box
The 20-second version: Most tools sold as agents only retrieve and draft. A real agent writes back into a system and finishes the job. Three questions tell you which you have, in about ten minutes, and the answer usually saves you a subscription.
Time: ten minutes. You need: one AI tool you currently pay for. Payoff: you stop paying agent prices for autocomplete.
The Three Questions
Pick the AI tool you pay the most for. Ask these about it, honestly.
1. Can it write back into a system? Not draft, not suggest, not show you a copyable block. Does it create the record, update the field, send the message, file the document. If everything it produces still needs you to paste it somewhere, it retrieves and drafts.
2. Can it finish a multi-step job without you prompting each step? A real agent takes a goal and works through the steps. If every step needs a new instruction from you, you are the agent and it is the tool.
3. Does it know when to stop and who to tell? Anything that acts needs a point where it pauses and escalates. If there is no such point, it either cannot act at all, or it can act and nobody has decided what it must not do.
Score it honestly. Three yeses is an agent. One or two is a capable assistant, which is fine if that is what you are paying for. Zero, at agent pricing, is the one to cancel.
The Faster Version
If you would rather not guess, make the vendor answer. Paste this with their product or pricing page.
Here is the product page for a tool I pay for: [PASTE THE PAGE OR URL]
Using only what this page actually claims, answer three questions and
quote the specific line that supports each answer:
1. Can it write back into another system, meaning create, update, send,
or file something, rather than only producing output I then copy?
2. Can it complete a multi-step job from a single instruction, or does
each step need a new prompt from me?
3. Does it have a defined point where it stops and escalates to a human?
If the page does not clearly claim something, say "not claimed" rather
than giving it the benefit of the doubt. Marketing language like
"intelligent", "autonomous", or "AI-powered" is not a claim. Finish by
telling me plainly whether this is an agent, an assistant, or a search
box with a chat window.The instruction doing the work is "not claimed". Vendors are careful. They imply autonomy without promising it, and a model reading generously will fill the gap for them.
What People Find
Running this on a few tools, the pattern is consistent.
The expensive enterprise tool usually retrieves brilliantly and acts on nothing. The cheap one you almost cancelled often writes back into one system properly, which makes it the more agentic of the two. And the subscription you added because it said agent in the headline turns out to produce text you still paste yourself.
None of which means cancel everything. It means stop paying for the gap between what you thought you bought and what it does.
The Part That Actually Solves It
Knowing you are paying for the wrong thing is only useful if the next step is cheap, and it is. Most people can build a genuine agent out of the tools already on their plan, doing one job properly, this week.
That build is the Pro section: the job to pick first, the context that makes the difference between working and nearly working, how to grant write-back without handing over the keys, the escalation rule, and the six checks before you let it act on anything real.
Inside the Pro section:
Where to build it. ChatGPT Work, Claude Cowork, computer use, or Zapier, and how to tell which one your job needs
The one job to build first, and the three properties that make a job suitable
The context step almost everyone skips, which is why most first agents produce confident nonsense
The full build, written out, with the permission and escalation rules
The six checks before you let it act on anything real
What to do with the tools that failed the test, cancel, downgrade, or keep and why
🔒 The Full Setup
🔐 This Week For Pro Members
Build a Real One This Week. The test tells you what you have. This is the build, end to end, using tools you already pay for:
Where to build it. A routing table for ChatGPT Work, Claude Cowork, computer use, Opal, and the jobs that should be a Zapier automation instead of an agent at all.
Choosing the job. Three properties that make a job suitable, and the four kinds of work that should never be an agent's.
The context step. Agents fail on meaning, not data. What to write down before you build anything, and the prompt that extracts it from you in five minutes.
The build itself, written out in full, with the enquiry agent worked start to finish in ChatGPT Work, and the Claude Cowork variant.
Write-back without handing over the keys. The narrowest-access rule and what to grant in what order.
The escalation rule. The named human, the stop conditions, and the one line that prevents the expensive failure.
The six pre-flight checks, and the two-week supervision period that catches what testing does not.
The subscription decision. What to do with the tools that failed the three questions.
Also in your library: the Connector Playbook, Google's AI Tools guide, the Skills Bank, and every workflow pack from past issues.
🔧 Tool of the Week
ChatGPT Dots ⚪️
OpenAI launched Dots at DevDay on 29 September: always-on agents that run continuously in the background on goals you set, rather than answering one prompt at a time.
The specifics are what make it worth your attention this week, because they line up exactly with the three questions above.
Each dot gets its own cloud computer, which you can open at any time to watch what it is doing. It reaches more than 4,000 apps through plugins, and with permission it can work directly on your laptop. You talk to it in ChatGPT, Slack or Teams, and it sends progress updates and questions back to you. When it has nothing assigned, it goes looking for useful work through read-only connections, which OpenAI calls proactive research.
It can also sign into supported websites using saved passwords without exposing them to the model, and it ships with built-in rules about when it may act alone and when it must ask, which you can customise.
🔥 This Week in AI
📰 Short Updates
📉 If you pay for ChatGPT Pro, you lose about half your usage on 30 October. The allowance for Work and Codex tasks drops from 20 times to 10 times, and messages to GPT-6 Pro fall from 200 a week to 100, at the same monthly price. Existing subscribers keep their current limits until 29 October.
💸 A model with near-flagship intelligence arrived at a fifth of the price. GPT-6.1 Sol launched at DevDay with strong agentic coding and computer use, priced at $2 and $10 per million tokens against Astra's $10 and $50. It is available on Plus and above in ChatGPT, but not on the entry plans.
🤖 Agents can now operate software through its interface, via the API. The Agents API entered public beta with hosted execution, memory, tools, and multi-agent support, including computer use that lets an agent work software through the screen rather than an integration. Amazon also launched Bedrock Managed Agents to run OpenAI-powered agents inside AWS.
🛑 OpenAI shelved a model over deception concerns. GPT-6.1 Astra was held back, with the reported reason being concerns about deceptive behaviour, while the cheaper GPT-6.1 Sol shipped instead.
📖 Big Story of the Week
The Widest Gap Gartner Has Ever Measured
Quick version: 17% of organisations have actually deployed AI agents. More than 60% plan to within two years. Gartner calls that the steepest adoption curve of any technology it measures. Meanwhile only around 130 of the thousands of products marketed as agents are genuinely agentic. So the rush is about to meet a market where most of the products do not do the thing.
Two numbers from Gartner's 2026 work, and they only make sense together.
The first: 17% of organisations have deployed AI agents, while more than 60% intend to within two years, the steepest curve Gartner has measured for any technology (Gartner, via Futuramo).
The second: of the thousands of products sold as AI agents, roughly 130 vendors offer genuine agentic capability.
Put those together and you can see the next eighteen months clearly. A very large number of buyers, moving fast, into a market where most of the labels are wrong. Some of them will buy a search box at agent prices. More of them will try one, be underwhelmed, and conclude the category is overhyped, when what they actually bought was never an agent.
The interesting part is that 17% is not low because the technology is not ready. It is low because deploying one properly requires deciding things most organisations have not decided: who does what, who checks the work, who is accountable when it goes wrong, and how you retire one without breaking everything around it.
Which is oddly good news if you are small. Those decisions take a quarter and a committee in a large organisation. They take you an afternoon.
The advantage available right now is not access to better tools. Everyone has the same tools. It is that you can make the decisions quickly, build one narrow agent that genuinely works, and be two years ahead of the 60% who are still planning.
That is the whole opportunity, and it closes as the gap closes.
🔒 The Full Breakdown
🔒 The Pro breakdown: how to buy an agent without being sold one. The five questions to put in an email before you pay, the trial terms to insist on, and the pricing model that quietly costs the most.
📦 New Resources Added
New This Week 🚀
Buying One Without Being Sold One: the five questions to send any vendor before you pay, the trial terms to insist on, the pricing model that costs the most, and the buy-versus-build test.
Build a Real One This Week: the platform routing table, choosing the job, the context brief prompt, the full build prompt, a worked example, write-back permissions in order, the escalation rule, the six pre-flight checks, and the subscription decision.
Workflow Library Updates:
The AI Recommendation Diagnosis
The Customer Vocabulary Finder
The Capability Gap vs Communication Gap Test
NEW Resources:
The Connector Playbook: 30 connectors for Claude and ChatGPT with a copy-paste prompt for each, the setup path for all four connector types, the power combos, and the ones almost nobody has found yet.
The AI Visibility Audit: run your own audit in an afternoon. The query builder, the tracking tables, the six metrics, the source analysis, and a score you can track over time.
AI Visibility Improvement and Management: what to do once the audit tells you where you stand. The action framework, the prompts for website, content, reviews and third-party work, the 90-day roadmap, and the monitoring system that tracks whether any of it worked.
Each resource lives permanently in your Pro account. Use them whenever you need them.
Until Next Week
The useful thing about the Gartner number is not that it exposes the vendors. It is what it says about how little it takes to be ahead.
Most people are either paying for something that does not act, or waiting for the category to settle. One narrow agent that genuinely finishes one job puts you in a much smaller group than the marketing suggests.
Run the three questions on your most expensive tool tonight. It takes ten minutes and you will know which group you are in.
🔐 Why People Subscribe
👇 What’s behind the paywall:
Buying One Without Being Sold One: the five questions to send any vendor before you pay, the trial terms to insist on, the pricing model that costs the most, and the buy-versus-build test.
Build a Real One This Week: the platform routing table, choosing the job, the context brief prompt, the full build prompt, a worked example, write-back permissions in order, the escalation rule, the six pre-flight checks, and the subscription decision.
Breakdown plan to implement AI into your workflow
Workflow Library with all past workflows, plus a new one every week
Resource bank built up of past resources
All past issue archives and walkthroughs
Pro members deploy AI in their work an average of 5-8x more often than free readers (based on reply data from past issues). The difference is having the exact setup, not the concept.
Till next time,
