When did Microsoft ever respect the customer?

What you're basically saying is, that you need documentation for AI to be good at maintaining it. If you already use AI or have documentation you're fine. Instruct the AI to keep it up to date. Otherwise maybe ask the AI to document the status quo and possibly add a plan for refactoring. Current AI is limited by context, so it is very good to keep less verbose information than you code base. Have understandable documentation about the how and why and it does not only help AI.

They can be VERY verbose to read. But ... we have AIs to summarize them, or to answer targeted questions. Let the agent read 100 plans and then explain how and why the feature evolved.

It was hard to write, so it should be hard to read!

Tell teachers who grade lyric interpretation. "That's not how the author meant it!" Sure?

It at least excels at the "How". Have you tried one of the IDE integrated agents in "ask" mode? You ask "How and when is X called and what are the side effects", allow it to read all workspace files without asking and let it work. When you come back five minutes later, you get a free text description, a table and some flowcharts with the answers. Of course you can also ask follow up questions.

They why may or may not work, just as for humans. If the code base is ... lets say suboptimal, then anyone reading it can only speculate on the why. You know, code quality is measured in WTF per minute.

Anthropic wasn't quite the charity before. Highest token prices, a lot of fearmongering, etc.
If you're looking for more user oriented AI look into the Asian direction. Mostly China, but LG's models are also nice.

"So? Again, why do this now? What is it about AI agents, particularly since agents aren't special, as you say?"
Because they seem to need the feature now. It also sounds like it reduces the overhead in discovery, which is good both for the agent and for the site that would need to be explored by the agent otherwise.

Creator: Eclipse License

I also saw (mostly in Vibe coded projects): "This is GPL, except you're making more than 1 million USD with it"
That's not how open licenses work. If you put an open license on the work, then your other restrictions are void. What you can do (with some licenses) is to modify the license. For example an adapted MIT license is possible. An adapted GPL is not. Using GPL and putting restrictions in the readme is as effective as objecting the facebook ToS with a post on your facebook wall.

Who says only AI is using agents? You're using a web user-agent right now. You probably already used a mail-user-agent, which delivered a mail using a mail-transfer-agent to a mail-delivery-agent. Don't be too impressed when someone uses a technical word that some people may associate with secret agents or whatever.

And they now just can lookup what's available on a site. From TFA:
> Thereafter, attempts to resolve the agent record _{agent-name}._{protocol}._agents.{your-domain} will return the relevant details, which can then be verified via DNSSEC, optional JWS signatures, and DANE policy before agents connect to one another.

That doesn't look too different from other SRV records.

It's basically SRV-Records for AI agents. Nothing too fancy.

10x dev/engineer is just bullshit. And AI estimates are rough, but we can probably be happy when they err on the side of caution, because our managers will use them.

Indeed, every good license contains such a clause. But this clause only works when the code accidentally does damage, not when it does so maliciously. I think you could try clever traps for AI agents which require a court case to determine if they are traps, but this one is obvious.

I can't confirm that. It rather takes a best-practice path with code that looks like from a tutorial, so one needs to cut a lot of stuff that may be useful but is not needed. LLM will (unless instructed otherwise) always add a lot of precondition checks, error handling, etc. that one might want to cut for simpler projects or when one knows that the precondition will always be met.

They also comment a lot and not just "increment i by 1" but really useful comments, but also leading to verbose code that looks like from a tutorial explaining the algorithm to the reader. This is good if you expect a novice to read the code later on, but may be wordy for people who know what algorithm it is.

I always find it funny when I start in Ask-Mode with "How complicated would it be to add feature X" and the LLM drafts some parts of the feature and then estimates like 10 human-hours, when I know that when I switch to code mode I will implement it in half an hour together with the agent. The numbers may also have been a bit too high for coding it myself (but better not underestimate effort!), but certainly the estimate the agent did does not apply to the agent itself.