Personal Blog where I write about things I learn or discover.
An agent harness is the software wrapped around a language model that turns it into an agent: the loop that calls the model repeatedly, the tools it can execute, the context management that decides what the model sees, and the guardrails that decide what it’s allowed to do. The model predicts text. The harness is everything that makes those predictions add up to work getting done. Anthropic’s…
On July 23, 2026, I opened an empty repository and started building Minute, a private meeting notetaker for macOS. Five days and 124 commits later, version 0.7.0 was available as a private beta for Apple Silicon and Intel Macs. I built the application through Codex. I set the product direction, chose the constraints, reviewed the work, and decided what counted as done. Codex inspected the…
You push a new image, update an ECS service, and the console says Deployment in progress . Then it sits there. A new task appears in PENDING . It changes to RUNNING , but the old task does not go away. A minute later the old one starts draining. Eventually it disappears and the service returns to steady state. If you only watch the task count, the whole thing looks oddly slow and a little random.…
Every few months I have the same conversation. A small team, three to eight engineers, is containerizing their app, and someone says “we should use Kubernetes, that’s the industry standard.” Six months later they’re maintaining a small distributed systems platform on the side, and the app they were supposed to ship is still competing for attention with CNI upgrades. I’ve written before about the…
Every few weeks someone downloads the GGUF build and the MLX build of the same model, runs both, screenshots the tokens-per-second counter, and posts it as proof that one format wins. The replies split down the middle. Half the thread says MLX is obviously faster, the other half says the test was rigged. They are both right, which is the problem. The number on the screen is real and it is also not…
Every coding agent session starts from zero. The agent doesn’t know how your code is organized, which files matter, how the pieces connect. It has to rediscover the architecture from scratch. Grep around, read some files, build a mental model, start working. That mental model disappears the moment the session ends. I kept watching this happen. Ten minutes of exploration before any real work, every…
I have too many issues and not enough time. Same as everyone. The usual loop is: pick an issue, context switch into it, write the code, open a PR, pick the next one. Do that until the sprint ends or you lose the will. Coding agents help with this. I can point Claude Code at an issue and let it work while I do something else. But that’s still one agent, one issue, one terminal. If I have 10 issues…
I’ve written before about how AI agents are just CI pipelines with an LLM plugged in . That post mapped agent concepts to infrastructure patterns you already know. But there’s a discipline forming around the infrastructure side of agents that deserves its own name. Harness engineering. It’s the practice of building everything around the LLM — the execution environment, tool definitions, safety…
Here’s something that frustrates me about coding agents. They forget everything. Every single session starts from scratch. The agent that spent 45 minutes yesterday figuring out your authentication flow? Gone. The decision to use JWT over sessions? Gone. The bug it found in your ORM’s lazy loading? Gone. You start a new session and it re-discovers the same patterns. Repeats the same mistakes. Asks…
In this post, I’ll show you how to think about AI agents through the infrastructure patterns you already use. Think about your CI runner. It spins up an environment. Runs some steps. Reads files. Runs tests. Captures output. Decides what to do next. Knows when to stop. Now swap out the hardcoded logic for an LLM. That’s it. That’s an AI agent in simpler terms. The fancy demos want you to think…