Came across a fascinating post by @arscontexta today about using knowledge graphs to make AI agents smarter. It hit differently because it described the exact problem I have been wrestling with in Iris.
The problem
Right now, Iris is like a chef who gets handed the entire cookbook every time someone orders a dish. Ask about Docker? She gets the Docker notes, but also the cooking recipes, the workout plans, and everything else. There is no filter, no structure, no "hey, this topic is related to that topic."
Skills today are flat text blobs. Imagine writing everything you know on separate sticky notes and throwing them all on the floor. That is roughly how Iris receives context. No connections, no hierarchy, just a pile.
The idea
arscontexta's post showed how structuring knowledge as a graph - think a web of connected ideas rather than a flat list - dramatically improves how AI agents reason. The magic is not just in finding the right information, but in preserving the relationships between ideas.
Think of it like a library vs a mind map:
- A library stores books on shelves. You can find a book, but the shelf does not tell you which other books are related.
- A mind map connects ideas with lines. You can see that "Docker" connects to "Networking" connects to "Load Balancing." The connections carry meaning.
That is what a skill graph does. Each piece of knowledge is a node (a card), and the connections between them are edges (lines). When Iris needs to answer a question, she does not grab random cards - she starts at the most relevant card and follows the connections outward, gathering a coherent cluster of related knowledge.
Why this matters for Iris
This could be transformative. Instead of dumping everything into Iris's brain, we give her a structured map of knowledge that she navigates intelligently:
- Ask about Kubernetes? She gets Kubernetes skills, plus connected Docker and networking knowledge, plus your personal preferences for deployment tools.
- Ask about meal planning? She gets nutrition knowledge, plus your dietary preferences, plus your shopping workflow.
The connections do the heavy lifting. No manual curation needed for every question - the graph structure naturally surfaces related context.
Initial sketch
Spent the evening sketching out the concept. Did not write a single line of code - that is what Claude and Codex are for. My job is the vision and the steering. Their job is turning the napkin sketch into production code.
The rough shape:
- Nodes are units of knowledge: skills, personal facts, workflows, references
- Edges are the connections between them (created automatically from links in the text, like how wiki pages link to each other)
- At query time, find the best starting nodes, then walk the graph outward to gather connected context
- Inject the result into Iris's instructions - she sees only what is relevant
- Keep it on a budget so we do not overwhelm her with too much context
Tomorrow I will sit down with Claude and start on the architecture.