Towards Building Agentic OSINT Analysts
Curiouser and Curiouser
Hi everyone. I’m really writing this article for myself today, to commemorate something, and you are welcome to read along. Today, I built the beginnings of some very powerful OSINT AI capabilities.
What’s OSINT? Enroll in OSINT for Everyone!
On Friday, I will be giving some Agentic training, so I am building out an Agentic OSINT workflow right now. You can just follow along with this article. I want to commemorate this important moment, so I am using this blog for screenshots and to record my thoughts.
I built World AI in 2025. I opened up API access a while ago. I added MCP last week. I am using MCP in this demonstration. I’m not going to show you how to build this, other than the high level function names.
The process is simple. I write as detailed of a prompt as I want describing the research I want to do. The prompt is then converted into a research plan.
That research plan is then converted to a data collection plan. I will extend this to include all kinds of data, and it currently includes Verdant Eye (high level world awareness) and GrooveSeeker (street level event intelligence). In this simple example, I just want to know what happened in the last two days.
My natural language endpoints are this easy to use. Your agents can send them simple questions. They will receive data back if they do. This article is a demonstration of that.
I then download the necessary datasets for the investigation. This is not web search. This is what completeness of information looks like. This is what my web crawlers saw on the internet in the last two days, and they are tapped into the MOTION of the internet itself.
Your AI agents and workflows could be world-informed if they tapped into my technologies. I provide world awareness that is useful in preparedness. I would rather know what is headed my way than be surprised.
I want to avoid dangers that are heading my way. I want to stay away from violence and crime.
I want to capture opportunities that are within my reach. I want to go see my favorite bands and go to free events.
It is not ideal to be slow at either of those two things.
I then denoise my datasets and rank the clusters.
Clustering gave me the main topics of the dataset. After that, I enrich the data, pulling in full text, images, links, and all kinds of useful metadata that my agents will use to answer questions. It is useful to know how to enrich data. This blog is about enriching data.
I then build a chatter timeline.
Which can be visualized.
I can inspect the clusters.
It can build out the information source landscape, identifying the top sources in the dataset.
Which can be visualized. This isn’t just for agents. I collaborate with AI. It does not replace me.
I can explore the information sources. I can see where the information came from and what the ecosystems look like.
I then extract evidence, which will be used to answer the question. And then I use the information to answer the question. But this is COMPLETENESS of information, not lol ten rows of web search data. Do you see that 622?
And then the AI answers based on clusters of evidence. So, there’s not one answer; there is a big complete answer. I asked what happened in the last two days and I want to know.
The answer is 813 lines long and provides source information.
See. It’s complete. It is not just one source or one single link. This uses evidence and different analysis techniques to help answer questions. This doesn’t just read documents. Haha. No way…
“But David! What if there is disinformation?!”
Good question! Good thing that’s a topic I understand very well. I am building in all kinds of cool language capabilities for narrative analysis, manipulation detection, sensationalism, disinformation tactics, etc. Yes, I do know this topic. I have been doing World AI stuff since about 2020. I have been writing about Language Models for a long time. I am not new here. My book is about Natural Language Processing and Network Science.
So, anyway. Yes, this will very soon be extended for narrative analysis, manipulation detection, sentiment analysis, and all kinds of cool things. But this is the 0.1.0. :)
That’s that! I started with one simple question: “what happened in the last two days?”
And I ended up with a 813 line long OSINT report about what happened in the last two days.
Oh, wanna see a fun one? I tested it earlier with some UFO stuff. :)
Pretty kewl! Don’t know why people aren’t climbing over themselves to get access. This is super cool and useful for staying informed and being prepared for anything. You can use this to research anything. Why aren’t more people curious about things? That’s weird to me.
Nice and easy, and makes sense. When I teach OSINT for Everyone, I explain how to do this. Now I have automation that can do it for me.
You can literally use this to study Artificial Intelligence. You can use World AI to study anything. You can use it to build new products. You can use it to create new companies. If you can use Google, you can use this.
This is extra cool, though, because it doesn’t give you web search results; it can do full OSINT on auto-pilot. This is way more important than a search engine.
And next I’m going to attach OSINT agents to this. :)
Hehe. Just kidding. Not those ones.
Just two weeks ago I taught a class called OSINT for Everyone. (Enroll in the next one!)
And tonight I have a fully automated OSINT analyst that researches whatever I want and gives me all kinds of cool data artifacts (graphs and datasets) to explore at my leisure. Super fun.
What would you build with this? I come from cybersecurity. I have a million ideas. Some are serious, and some are really fun like GrooveSeeker.
Speaking of which, GrooveSeeker is out now for iPhone. It is free and it has no ads. You should download it and use it. It’s very cool. And then tell your friends. It is already global. Help me get a hundred million users. :)
Thanks for reading this article. I automated this dude today and I’m really happy about this. My World AI and Verdant Eye MCP is extremely cool and useful. Can’t wait to get my agents hooked to this. Should have that working tomorrow. I move fast.
Please subscribe and tell your friends about this blog.