Agentic AI writings
I love getting to share counterintuitive ideas about building agentic AI with other people, and I mostly stick to Twitter when I write about it. Here’s a small collection of my writings, but you can find a lot more if you follow me on Twitter.
Traps in building agents
This thread is about some seductive ideas that sound very smart but are actually a massive waste of time when building agentic AI. You can read through a large discussion that happened on Twitter about this.
In building AI agents @cline , we've identified three mind viruses Mind Viruses are seductive ideas that sound smart, but don’t work in practice.
— Ara (@arafatkatze) August 19, 2025
1. Multi-Agent Orchestration
2. RAG (Retrieval Augmented Generation)
3. More Instructions = Better Results
Let's explore why! pic.twitter.com/p2p93Dt9QM
Agentic search
Agentic search is a very difficult unsolved problem and has historically had all different kinds of strategies. The most common ones are grep and vector search, but each has its pitfalls. It’s still unclear to me what the actual answers are here.
Cursor vs. Cognition have opposite takes on agent search:
— Ara (@arafatkatze) November 12, 2025
- Cursor: Custom embeddings trained on agent traces improve accuracy by 12.5%
- Cognition: Embeddings are counterproductive, so we trained models to use grep with 8x parallel tool calls.
Both have benchmarks.
Hmmm.. https://t.co/QOKy6kvSVA pic.twitter.com/OhSKOTFMLw
Self hosting inference mini book
A mini book on self-hosting inference, covering the practical details of running your own inference stack.
— Ara (@arafatkatze) August 4, 2026
Recursive self improvement
A striking example of recursive self improvement in action: an agent improving its own harness to boost its performance and cut its run cost.
We had Kimi K3 recursively self-improve the Cline harness to improve its own performance.
— Cline (@cline) July 29, 2026
17 hours later, it went from 77.5% to 88.8% on Terminal Bench, and cut run cost from $79 to $49.8. pic.twitter.com/7eq9FjeUdu
Open source model labs
This post is about how open source model labs like MiniMax are actually doing extraordinary work with smaller models, which are just as good as the larger ones.
Many SOTA models have over a trillion parameters, but this only only has 10 Billion AAAAAND its open source.
— Ara (@arafatkatze) December 23, 2025
Hmm..... https://t.co/5J47EabcSf
Doing things and telling people
Many of us are just lurking through the internet and not sharing, thinking our ideas are not worth sharing, and this post explains why sharing ideas is critical not just for yourself but for the world.
Hey @swyx question for you,
— Ara (@arafatkatze) December 23, 2025
You always say "do things, tell people." But I find the doing interferes with the telling.
I get lost in the doing. I have to forcefully block time on my calendar just to force myself to tell people by writing. And even then it feels like I'm spilling…