Agent-errific: Everyone’s Talking About Agents and the Agentic Web - But What Actually Are Agents?
Agents, The Agentic Web, It's Agent-errific
There's lots of discussions at the moment around "the agentic web", giving the impression that autonomous "agents" are going to intelligently do all kinds of smart things for us:
"The BluePrint: Web 4.0 - The Rise of the Agentic Web". This has a nice definition of digital AI agents as: "AI systems that are designed to perform digital tasks by interacting with digital ecosystems either under human guidance or autonomously" and has a good framework in which to understand AI agents. Nice article.
Microsoft in particular seem to be believing heavily in all things agentic: "Microsoft Build 2025: The age of AI agents and building the open agentic web"
And Agentic AI cropped up at a recent AI Summit in London, UK: "AI Summit: The Agentic Leap: From AI That Knows to AI That Does" & "AI Summit: Agentic AI: Systems That Think and Act for Themselves"
But what exactly is an agent? Is it more than a process that sits quietly running by itself, sometimes querying some GenAI/LLM, and then processing the response, and perhaps reacting to some inputs from elsewhere? If so, in what way is it more?
Let's Try to Build an Agent
Having successfully experimented with ThreeJS & gained confidence that my coding skills were still reasonably functional (see ThreeJS Acid House experiment), and got used to working in partnership with ChatGPT (see Using ChatGPT post), I figured the best to find out more about "Agents" was to try to actually build one, so, with ChatGPT's help I'm going to try to build an agent using:
What Am I Going to Try to Build?
Something that has grabbed my attention a bit is the subject of "heat networks" - basically the use of hot water in old coal mines to heat houses and businesses. So, here's the brief of what I'm going to try to build:
- MVP: An agent that generates a daily feed of news articles around "heat networks". That's it. Nice and simple.
- Nice to haves:
- (a) The daily feed only contains articles not previously included
- (b) The agent provides some mechanism for listing all articles
- © The agent provides some mechanism for searching articles
- (d) The agent provides some kind of totally unnecessary ThreeJS 3D experience
- (e) The agent is actually doing something that only an agent could do. No idea yet what that would be. Might not get to this stage.
Steps Trying to Build This:
Learning: 4/7/25:
- ChatGPT is excellent at helping with coding syntax (caveat below)
- However, ChatGPT is absolutely terrible at helping navigate organisations help docs, APIs, etc - particularly if you're working with Beta versions. It's so, so, bad. It will just make stuff up that wastes hours and hours of your time. You cannot rely upon it when doing this.
- So, ChatGPT is excellent at helping with coding syntax provided you point it to the correct and up-to-date sources.
Learning: 17/7/25:
So, two weeks later:
- Started building a Retrieval Augmented Generation agent using Google Vertex AI and Google's Agent Development Kit.
- Managed to get some PDFs uploaded to Google GCS & to rebuild a Google Matching Engine index based on that PDF data.
- However, when it came to using a MatchServiceClient to query the index and find matching chunks of data to then feed to the LLM, it eventually turned out that you need a private endpoint to do this query.
- And to use the private endpoint you then need a Virtual Private Cloud (VPC), at which point merely running and debugging was becoming a huge pain.
- And also using Google's Beta software with ChatGPT is pretty painful - it just kept making recommendations that were out of date.
- So, back to the drawing board. And try to find a more simple way to do this.
- A small side note: at the moment Google's Agent Development Kit appears to be a pretty simple wrapper around an LLM. Not that impressed really.
Overall:
- Another week passed after the previous update, not that productively.
- Overall, after spending a few weeks on this - not a wildly positive experience.
- Google's software is still in Alpha and Beta and as at the time of writing, somewhat painful to use.
- ChatGPT is not that helpful when working with things that are relatively brand new e.g. Alpha and Beta level codebases.
- The Google's Agent Development Kit was pretty underwhelming at the time, and felt like a thin wrapper around a LLM.
- The whole Google Vertex AI felt like using a massive hammer to hit a tiny nail.
- The whole idea of an agent using an LLM felt scary and unpredictable: e.g. "do this thing please within these parameters" - and it might/might not do what you ask - which isn't ideal really.
- So to answer the question of this post: "What are agents?" - Agents that are primarily LLM-based frankly, feel like code that is putting an LLM in a very tight straight-jacket and hoping that it will do what it is asked. I suspect that with the very latest LLM models and a large amount of code around an LLM keeping it focused it will be possible to do something useful, the question will be whether that effort is worth the value produced.
Next:
I think I'll try something with a simpler toolkit, perhaps something around Retrieval Augmented Generation