Agents

Almost a year ago I wrote about how to keep the discipline in AI-assisted TDD. In that flow I steered the agent in real time, in the same session where the code was being written: I picked every test, corrected the design on the fly and approved every commit. That flow changed gradually. Models got better from one version to the next, and the repository conventions got sharper with every correction I wrote down. With both, it made less and less sense to steer the implementation step by step, and more sense to put the weight on the spec: make clear what to build and how, and delegate the rest. At some point the implementation was running on autopilot until it was ready for review. ...

How you integrate an LLM into your application fundamentally depends on who has control: does your code decide the flow while the LLM just responds, or does the LLM decide what to do next? This distinction leads you to two very different architectural patterns. Pattern 1: LLM as Infrastructure When the interaction is simple—you send a prompt, you receive a response—the LLM is infrastructure. It’s just another external service, like a translation service or geocoding. ...