An agent is a loop: the model proposes an action, a tool executes it, the result comes back, and the model decides what to do next. Everything that makes agents useful — and everything that makes them fragile — lives in that loop.
Tools are the interface to reality
A model can only act through the tools you give it. Underspecified tools produce confidently wrong calls; over-specified ones produce paralysis. The craft is designing tools with clear contracts, honest error messages, and outputs the model can actually reason about.
Control the loop, not the model
Reliability comes from the scaffolding around the model, not from hoping the model behaves. Bounded retries, explicit stopping conditions, and validation between steps turn an unpredictable generator into a system you can reason about. The model is one component; the loop is the product.
Observability is non-negotiable
When an agent fails, it fails several steps deep, and the root cause is rarely the last action. You need traces: every prompt, every tool call, every result. Without them you are debugging by vibes, and agents punish that harder than any other system.