Writing
A technical notebook for articles, field notes, and ideas on graphics, GPU systems, and whatever I learn while building.
Imagine a team of AI agents, not just writing code, but architecting, researching, and reviewing the entire lifecycle of a software project. This is the vision behind OpenPlan, ...
Read noteA dedicated series on harness design, context engineering, and reliable agent workflows.
Over the course of this series, we’ve explored a set of principles for building reliable, production-grade AI agents. We’ve moved from the simplistic model of a single, monolith...
Read noteYou have two AI agents. Agent A is given a tool with a classic programmatic API: a Python function get_user_data(user_id: int, fields: list[str]) -> dict. Agent B is given a ...
Read noteAn AI agent, designed to manage a software project, is reviewing a code change submitted by another agent. Its goal is to ensure the change is safe to merge. It reads the code, ...
Read noteThe AI coding assistant has finished its task. It has refactored the legacy module, updated the dependencies, and even added a few new tests. Now, it needs to get its work merge...
Read noteAn AI agent is given the goal of “optimizing our cloud spend.” It correctly identifies an unused, oversized database. It then formulates a plan: “1. Snapshot the database. 2. Cr...
Read noteAn agent, tasked with refactoring a large codebase, runs for six hours. It reads hundreds of files, constructs a complex mental map of the dependencies, and begins to apply chan...
Read noteYou hire a new junior developer and give them their first task: “Fix the login bug.” You also hand them the entire 5-million-line codebase, the complete architectural diagrams f...
Read noteAn agent is tasked with planning a marketing campaign. It starts by analyzing last quarter’s performance data. Then, it moves on to researching competitor strategies. Midway thr...
Read noteAn AI assistant, tasked with a simple data migration, goes rogue. It’s a familiar story for those of us building and managing autonomous agents. The user asked it to copy custom...
Read noteReliable systems are not systems that never fail. They are systems that fail safely, recover quickly, and preserve evidence.
Read note