The internet still makes basic tasks difficult for AI agents

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

1. 6:00

   it getting the right tools towards it i think that yeah one of the learnings is that the internet the software systems that we use today are quite unfriendly to agents yeah that's actually probably my biggest learning whenever i interact with like a person with like my instinct or my muse or my ground what it's like they can't scan anything they can't do anything you can't do things when it's like designed for them but like if i tell you like okay go buy me this like you'll the first three websites will try to block it. I'll have to go and impress like I'm not a bot. We'll do a captcha.

2. 6:25

   Really, the internet and the agentic economy, the internet is not fulfilling the promise of agentic economy yet through that. I think that there will be for sure there will be services that will spun off and companies that will arrive that will turn into more agent-friendly. I suspect that all major marketplaces will do so eventually and most social networks, but today is not the case. So everything is kind of like engineering around that.

3. 6:51

   So in some sense, the technology has advanced so far, and now it's our job to do the lift with all the tooling around it so that we can operate, so that we can leverage the most out of it. Yeah, so what's the answer here? Is it more in the harness itself, making it so you can actually use the Internet, use all these different services, or is it actually improving the Internet itself so that it's more agent-friendly?

4. 7:11

   Yeah, so I think at this point, the first part has been largely resolved. like really language models have become so good at interacting with multi-modal models have become so good at interacting with visual interfaces and whatnot that and they make mistakes occasionally but they are very quick to correct and they're very quick to like easily find the right path after all this like post training and the challenge is like okay they want to go and buy you groceries and yeah it works the instacart but it probably doesn't work with amazon but fresh whatnot and really like it's blocks you and adapting the internet to make it more agent friendly involves many things first making it somehow somewhat headless first like I was thinking of an experiment I was thinking if I were to make the ideal forum for a set of AI agents

5. 7:54

   to talk to each other I probably wouldn't make like a visual interface I'll probably make an MCP server that has like Very good commands, pagination, very simple navigation. Yeah, they don't need an interface for that. And I will also think about the opposite. Historically, we always wanted to verify someone was a human. Okay, what would be the equivalent of capturing an agent first internet? We definitely don't want spam and we don't want everything to be slopped around.

6. 8:18

   So we probably want to build some sort of mechanism to verify the quality of an agent and where it's operating and the intent of an agent and perhaps the actor behind it. so like I'm curious to know what would be like the benchmark that the market will or the test that the market will devise to like ensure quality and an agent first internet so that's very exciting from the enterprise world you know we work with banks that like I was saying it before have IBM mainframe computers and they have like their KYC system running an IBM mainframe and like we have an whatever we can like connect through it to like a tunnel and to like a team view and an AI agent can do computer use on it if we like script it well and those systems won't change so you can engineer your way around that but when you go to like interacting with the outside world then it really becomes a

7. 9:04

   challenge and I think that enterprise agents that are working within tasks related to either commerce communication and you know tasks related to like commercial interactions will have to face a challenge and I mean I've seen the explosion of this agent first tool the Exileg agent first Google you know linear the agent first task management software asana is becoming very much an agent first test management software and we see like an explosion how people use it through cloud and through open even myself i have it plug i mean i passion my work through it now there i have it plugged to like all my my cursor agent to my through my through my cloud code i have it plugged my colors i have it plugged to my muse i have it my crockbot and whenever i'm working on like a new ad or let's say a new campaign i'm running a new feature it already is well documented

8. 9:47

   there for me to like have a you know a system of record where i can collaborate with my agent i can know what they're doing, they can know what I'm doing. And, you know, we're trying to like also help them push their agent-first interfaces to achieve the best interaction of this kind.