Much of Adam Engst’s skepticism of AI agents touches a sympathetic nerve with me. I don’t intend to have any AI agents handling calendar, mail, travel plans, bookings, task tracking, etc. Anything that requires understanding my internal context and intentions.
Katie
The analogy to paid help is a good one. After The Four-Hour Workweek there was increased interest in virtual assistants. For many tasks, they had the same issue. You need to invest a lot in the shared understanding of your life, goals domains of expertise and limits–as you’ve done with your research, for example. It’s similar with these non-deterministic automations we’re all looking at.
This is fine in terms of intent. An AI agent is proactive and normal AI is reactive, but they both use the same semantic database, which AI creates irrespective of your use if AI is on.
So if your device has AI all that data will be collected into a semantic database whether you use agents or not.
That post was spot on. I have found AI helpful for planning trips, but I would never hand it off to an AI agent to actually book the hotel or plane flights! I had it create a table comparing the different options for some flights and discussed it with my wife. There are two people involved in making this decision. How could AI possibly know enough about the both of us to autonomously make the right decision?
Unfortunately all of the hype around AI obscures what it is good at and makes it hard for people to use the tools like normal tools. They are good for some things, bad at others.
Here is another small example, I own a Tesla and I was driving along a two way road with autopilot engaged. We approached a strip of road where the lines were painted badly. The autopilot freaked out and jerked the steering wheel to the right, luckily I was just driving along at 30 mph and was able to correct it before it veered off the road. If AI isn’t smart enough to not be fooled by some badly painted lines on the road, how can you trust it to do all these other things autonomously?
All what data? Where? Sorry. I don’t have AI installed here that does what you claim.
Katie
What do you mean by “if your device has AI”? I have a local model installed on my Mac that I use with Claude Code for simple and straightforward tasks. It’s not going to turn anything into a semantic database unless I ask it to.
I asked Claude if it trains on our conversations. It said no but said its information may be out of date, want me to check. I said yes.
" The policy changed since I last had reliable info, so my earlier answer was out of date. Here’s the current picture:
As of the 2025 update, the default flipped. Anthropic may now use your chats and coding sessions to train models — but only if you allow it, and you control this in your account settings (Privacy → data usage). It’s an opt-out: if you don’t turn it on, your conversations aren’t used for training and are kept on a 30-day retention. If you do allow it, retention extends to five years."
we had a longer conversation and that may be wrong too.
Yes, Anthropic’s models will train on your data if you let them. (I’ve turned that off.) That’s not the same as creating a semantic database using your data for an agent to use, however.
The difference between an agent and general AI is in the proactive behaviour of the agent. The agent makes its own decisions based on the task — e.g. look through my email and calendar, remove all cancelled meetings, and add all the new ones. With general AI, you’d typically do this in stages: ask it to search your email, show you the relevant calendar entries, and you make the edits yourself. The difference is in the activity — how autonomously the AI acts. The data it works from is the same in both cases.
This means whether or not you use an agent makes no difference to the underlying data. So if AI is running on my Pixel and I want to search for photos with a beach and my wife in them, the AI needs a semantic index to fulfil that request, since it needs context beyond plain keywords or metadata. Crucially, that index isn’t built on the fly for my query — it’s already been built and maintained in the background (as photos are taken or synced), so my search just runs against existing infrastructure rather than triggering fresh indexing.
An agent set loose on the same device would draw on that same pre-built semantic index, plus the same raw data, to achieve its goal. The only thing that changes with an agent is how much decision-making it does on its own — not what data or indexes it has access to.
This is coming from the premise that you have AI toggled on on your device or given AI permission to access your data as in Cowork or Gemini on a Pixel.
What you’ve described appears to be the latest version of Siri AI—i.e., Apple’s specific implementation of a locally installed LLM, an agentic harness, an on-device vector database, and routing to a cloud model if the local model isn’t up to the task. That’s Apple’s architecture, but the only part of the equation that’s necessary to agentic AI is the harness. The on-device vector database—or access to any vector database, for that matter—certainly isn’t a requirement: it’s only necessary for a RAG (retrieval-augmented generation) based system. Siri AI requires RAG; a local model hosted on, say, Ollama and plugged into Claude Code or Codex does not.
The proactive/reactive framing isn’t wrong, but it’s incomplete and can be misleading. The core distinction between agentic AI and a chatbot is tool use and multi-step execution. A chatbot takes input, generates a response, and stops; each turn is independent and there is no persistent state.
An agentic system does one or more of the following: it uses tools, executes multi-step processes, maintains state across those steps, and operates with reduced human-in-the-loop. Note that the latter does not equal no-human-in-the-loop; it means that the agent doesn’t need approval for every step, not that it can act unprompted.
I gather that macOS Golden Gate and iOS 27 will automatically re-index Spotlight to enable semantic search via Siri AI, whether or not you intend to use it. I’ve got no problem with that, personally, but your mileage may vary.
That’s a pretty broad brush.
If you fly the same airline route regularly and want to check the fares daily/weekly then an agent might be ideal.
There are a million other examples.
Like any other software or any other tool, some are a good fit for your job and some are not.
But I would never, ever, contemplate getting the AI to book it for me.
It’s that final step: allowing an algorithm with delusions of grandeur, no sense of reality, and the propensity to make stuff up, to take possibly irrevocable action, with damaging consequences for which you, and only you, will bear the responsibility, seems to me to rather brave…[1]
I just don’t see any circumstances where this could be a sensible risk for anyone to take.
[1] For those who don’t speak British English, ‘brave’ here is roughly equivalent to ‘foolhardy to the point of recklessness, have you completely lost your marbles?’ cf Sir Humphrey Appleby, Yes, Prime Minister, BBC, London.
Most agents would not / could not even do that. An extremely common red-line for agents is to require explicit human intervention before making a purchase.
It might however be very convenient for an agent to fill in all the details subject to your subsequent editing and/or approval.
As I understand, the case for AI agents is precisely that they can execute commands without human interaction – the very word ‘agent’ means that they can act, and it’s what differentiates them from ‘ordinary’ models.
They may or may not be programmed with a promise that they will comply with red lines, but there are plenty of examples of AI ignoring instructions, or interpreting those instructions in a way that the human didn’t anticipate. Even if the guardrails were 100% effective – which doesn’t appear to be the case at the moment – would you be entirely confident that you can identify every single way in which the agent could misinterpret your red line?
Given that I don’t trust either myself to do that, or the model to comply with any restrictions, the risk seems far too high for the rather trivial benefits that are being touted.
In that case why use computers at all?
Are you 100% sure nobody can hack into your private photos/files/data and post them on the web or steal from you?
Or do you do your best to learn about a technology, weigh benefits and risks, and proceed accordingly?
There’s a bit of a large leap from ‘I don’t trust a specific new technology which has identifiable risks that have already caused problems’ to ‘why do you trust any technology at all?’ isn’t there?
There is always a weighing of risks, of course, and there always has been, just as there is whenever one does anything, such as driving a car.
At the moment, my assessment is that the risk of agentic Ai is towards the Should I drive 100 miles in my car with the dodgy brakes in dense fog to get to somewhere I don’t really want to go end of the risk scale. Some people will take that risk, either because they just like risk, or because they are desperate for the promised benefits.I won’t.
I’ll keep an eye on it, but unless it evolves to do something I actually want or need, then I’m happy letting those with higher risk thresholds walk through the minefield first.
The author of the article originally quoted has obviously not even done basic Google searches on the topic he writes about. AI agents are performing all of those tasks already.
Nothing to see here, keep moving…
Not to mention that he leaves out the the knowledge work-related tasks that AI agents do well! Things like extracting data from a year’s worth of account statements and turning it into a spreadsheet; or taming an out-of-control downloads folder; or, making order out of five pages of badly-scrawled meeting notes; etc etc etc.
The article is focussed on things like agents making restaurant reservations, booking hotels, buying event tickets and the like and not the things that take the drudgery out of knowledge work.