The attack was hiding in a social feed I run agents that read the internet for me. Every day they pull in posts from Reddit, X, RSS feeds and the web, and turn them into notes I can act on. For months...
The memory layer kept breaking. The wiki never did. For months I treated my AI agent’s memory as a stack of layers, and I assumed the smartest layer was the one doing the real work. That layer w...
Your AI’s memory grows — but does it ever forget? I run a fleet of AI agents, and for months I treated their memory like a bucket: the bigger the context window, the better. A million tokens? Gr...
The day I started questioning my own tool architecture I run a fleet of AI agents on a harness that gives them typed function calls — structured tools like read_file, search_files, and a dozen MCP ser...
I measured my AI’s tool usage. 90% of its work was done the hard way. Last week I ran a measurement on my own cron jobs that I’d been putting off for months. I have a fleet of scheduled ag...
Your AI’s memory grows. But does it ever die? Yesterday I wrote about how my agents act on timers instead of their own memory. Today I hit the next question, and it’s the uncomfortable one...
My agents are proactive. But they’re not self-driven. I run a small fleet of AI agents that message me on their own. Zora checks in a few times a day with something she found interesting. Charly...
The harness is replaceable. Your context layer is not. Last week I ran an eight-harness benchmark on my own model and my setup finished in the middle of the pack. My first instinct was doubt — did I b...
The 80% that hides the real problem Yesterday I published a post about my AI scoring 50% on a public benchmark — and how that single number hid the real story. Today I want to go one step further, bec...
My AI got a mediocre score in a public benchmark. Then I read the methodology. Last week a benchmark made the rounds claiming to compare eight different AI agent harnesses running the same model. My o...