MCP is far from perfect but if you just think of it as "LLM-optimized self describing 'stuff' for agents" it's a great set of primitives.

I think the consumption of MCP by clients should be way more sophisticated than the current SoTA. Something I'm hoping to explore more soon.

Jeff Bezos called it an "industrial bubble" and I agree.

There are tons of startups and maybe some big players that are going to get wiped out eventually, but the technology itself will be transformative and have enormous long-term economic value for those who figure it out.

The level of investment is substantially larger than either of the other bubbles - so if the potential is overestimated it's going to be very bad for lots of people.

But I'm bullish that AI is a bubble where the real potential is enormous, but the applications is uncertain.

Crypto was a bubble where the perceived value of the technology was massively inflated compared to the actual application value (stablecoins for money transfer and Bitcoin are the only notably useful crypto applications imo).

So the question becomes: is AI more Dotcom or crypto?

More detail including a code sample for trivially implementing linking in an agent on my blog: mbleigh.dev/posts/co...

Links were powerful enough to build the entire web...I think they'll be critical for building the context for agents as well.

I've been seeing more and more opportunties leverage links for context management that is both token-efficient and tool-efficient. We're using it in the Firebase MCP Server today to power our prompt-driven /firebase:init command!

We're *so close* to having this be widely available via MCP Resources, but MCP clients today either ignore resources entirely or only use them for @-mentions explicitly requested by the user.

We need to give models direct access to resource URIs to unleash the power of links!

You can build a single tool (let's call it `read_resources`) that can load context via URIs. These can include custom schemes like `support://some-topic` or even `user://self`.

Once you have this tool, models can bring in additional context on-demand so long as they see the URI.

REST purists long dreamed of APIs fully discoverable and navigable by hyperlinks. It was never practical. Even if machines could parse links, humans had to understand/navigate it.

LLMs change this. They are clients capable of navigating Hypertext as the Engine of *Agent* State.

Yes, like any other tool it needs to be used correctly and for things it is suited for to be effective.

Tools that help experts be more productive are pretty great tools!

Which makes me wonder if this is part of the difference between people who think coding agents are super helpful and those who think they're useless...do you accept them as an imperfect but nonetheless very useful tool, or do you treat them like they're supposed to "just work"?