Spider-Man Accidentally Explained the Hard Part of AI Product Design

Tony initially hides most of the suit behind the “Training Wheels Protocol.” Peter, naturally, gets Ned to disable it.

No training. No gradual onboarding. Just give me the full product.

Then Peter discovers that his web shooters have 576 possible configurations.

The suit has become dramatically more capable. Peter has become dramatically more confused.

For all his technical genius, Tony Stark apparently sucked at AI product design too. He over-engineered the hell out of it. Product managers everywhere can take some solace in that.

The problem is not that the configurations are useless. Any one of them might be exactly right in the right fight. The problem is that Peter has to search the suit's internal complexity while someone is attacking him. The interface has exposed the architecture instead of understanding the task.

The scene feels familiar. Models, agents, tools, MCP servers, memory, permissions, workflows, reasoning modes, autonomous execution.

Each capability is useful. Put enough of them in front of the user and you recreate the complexity AI was supposed to remove.

This is easy to miss because teams evaluate each addition separately. A model selector seems harmless. So does a tool picker, a permission dialog or an autonomy setting. The cognitive cost appears only when the user meets all of them at once.

Karen does not offer “Instant Kill” directly. She proposes Advanced Combat Mode, which sounds entirely reasonable for the occasion. Peter is in a fight. An advanced combat mode is exactly the sort of help he thinks he needs.

“Yeah, let's do it,” he says.

The suit activates Instant Kill.

Peter's response is immediate: “No, no, no, no.”

It may be the best enterprise-agent demo in the movie.

We talk about AI autonomy as though more is obviously better. But the technology can move much faster than user trust.

The cost of a bad recommendation is that I ignore it. The cost of a bad autonomous action is that I may have to undo what it did to my business.

The important variable is not autonomy by itself. It is the blast radius of a mistake. Let an agent rename a draft or summarize a document with little ceremony. Reallocating inventory across 600 stores deserves a preview, an explanation and a clear approval boundary.

For consequential work, the path probably looks like this:

The model may be capable of jumping to the end. The user may still need to travel through the intermediate steps.

Later, Peter is stuck in a warehouse. He starts talking to Karen to figure out how to escape. That is a perfectly functional use of an assistant.

Then the conversation wanders to Liz, his crush. Eventually Peter realizes they have been talking about her for 37 minutes.

He was supposed to be escaping from a warehouse.

In 2017, this was funny science fiction. In 2026, it is opening ChatGPT or Claude at 11 p.m. to think through one work problem and resurfacing an hour later discussing your career, relationships and life choices.

Traditional software tells you what it is for. You do not tell Excel about your marriage or ask Photoshop whether you should change jobs.

A general-purpose conversation has almost no such boundary. The same box can analyze a spreadsheet and discuss your ambitions. So people use it for both.

That openness is part of the product's value, but it changes the design problem. Purpose, memory and privacy can no longer be inferred from the screen. The product has to make them explicit.