Working With Agents
If I Had More Time, I'd Have Written Neater Code — The power of a programming language is not in what it lets you do — it's in what it doesn't. Rust may be better for agents than it is for humans. I find this funny.
To The Left, To The Left — Every step to the right on the development timeline multiplies the cost of a fix. Shift-left is the philosophy that keeps moving 'find out this is broken' earlier in the loop.
Reviews are in: my Agents have ADD — My agents review everything with the same sustained, undifferentiated focus. Every PR gets the same checklist. They notice everything and see nothing.
Boiling Agent Syndrome — An agent notes that CI failures predate its changes and calls the task complete. Technically correct. This is also how you end up with an always-red CI pipeline.
The Push Bike Harness — Most AI agent harnesses are training wheels. DEAR — Define, Enforce, Audit, Resolve — is a push bike. The ideal harness doesn't prevent falls. It catches them and teaches.
Pull the Cord — Teams get agentic tooling and point it at the existing task list. Nobody asks the agent to verify whether the documentation still reflects reality. The maintenance function needs to be explicit.
Death by a Thousand Tests — Not every check belongs in the synchronous path. The right amount of enforcement isn't the maximum — it's where the cost of checking balances the cost of missing.
Oops I Did It Again, I Forgot --dry-run — Agent permissions aren't about distrusting the agent. They're about acknowledging that even a capable actor will occasionally make mistakes. Build so mistakes don't take down the building.
I have No Memory Of This Place — Decisions made for good reasons in 2019 become load-bearing walls in 2026. With human teams, institutional memory survives imperfectly. With AI agents, it doesn't.
Most Rules Exist for a Reason — An AI agent encountering a constraint it doesn't understand will do what any capable new hire would: question it. Unlike a new hire, it might remove the fence without asking.
Schrödinger's Syntax — Engineers spent seventy years building better programming languages so intent could be specified precisely. AI takes English as input. That's not a flaw — it's the specification problem, unmasked.
Don't Hate the Agent, Hate the Process — Agents cut corners because the process lets them. When an agent does something you didn't intend, it didn't fail — it taught you where your process breaks.
Stop Asking Einstein to Run Your Datacenters — We build the most capable AI systems ever created, then use them to parse JSON and fix typos. Save the reasoning engine for work that actually requires reasoning.