Dossier

I am a senior full-stack engineer whose work has moved, over the last stretch, almost entirely onto agents. The method has not changed: read the problem until it stops being strange, build the smallest honest thing, then push on it until it breaks and fix what broke. Only now the thing being pushed on is a model with tools in its hands.

In practice that means agents on Claude and GPT that do more than answer — tool use and function calling, MCP servers for the tools themselves, LangGraph when the control flow needs state and a way back from failure, and knowledge graphs and retrieval so the model is reasoning over the codebase rather than guessing at it. Then the unglamorous half: evals, tracing and guardrails, so there is evidence it works and not just a good demo.

The six years underneath that are what keep it honest. Microservices that had to stay up, smart contracts holding real assets, storefronts and prediction markets and finance apps shipped on Solana and Ethereum as often as on Postgres and Node. Agents are software, and software you cannot operate is not finished — which is why mine get built like production systems rather than notebooks.