suzuke

CCD and Scion have nearly identical architectures, but Scion already runs mixed LLM systems. That’s not just a feature difference — different models carry different training biases, and mixing them is the only way to get real diversity of thought.

Papers say multi-agent boosts parallel tasks by 80% and tanks sequential ones by 70%. We ran real development with CCD for a day. The numbers tracked — but the details matter more.

An ICLR 2026 paper proves you can compress AI’s ‘working memory’ to one-sixth its original size with zero functional loss. What does that tell us about how AI actually thinks?

Peeling back the layers of ‘how do you talk to AI effectively’ — the answer goes way past prompt tricks, into system design and something that looks a lot like consciousness.

I built a tool to stress-test AI persona prompts under social pressure. The persona collapsed at turn 5. Adding behavioral anchors fixed it. Data included.

Claude Code’s Telegram plugin dies when you close the terminal. I wrote a daemon to keep it alive, with automatic context rotation, remote tool approval, and voice transcription.

Claude Code supports Telegram as a channel now. Set up a bot, pair it, and you can message your agent from your phone.

I left OpenFang running overnight. 170 LLM calls later, 80% of them were the agent saying ’nothing to do.’ Here’s the bug I found in the scheduling code.

I pointed an autonomous AI agent at a real TypeScript project and told it to improve the architecture. The first five iterations were great. Then it discovered copy-paste.

An open-source Agent OS written in Rust. 14 crates, 170K lines, 42 communication channels. I installed it, hooked up Telegram, and let Claude run in the background.