Philipp Schmid says agent engineering requires a different mental model from conventional application development. Instead of prescribing every step, engineers define the destination and let the agent choose a route, then observe behavior and improve the surrounding instructions, tools and checks.
He identifies five practical differences. Text becomes a primary form of state, conversation can change the plan, errors should be returned to the agent as inputs for recovery, reliability must be measured through evaluations rather than only unit tests, and tool APIs need enough semantic context for a model to use them safely.
Schmid recommends building agent systems with the expectation that some custom logic will be deleted as models improve. The durable engineering work is the definition of intent, the quality of interfaces, the trace of what happened and the evaluation of whether the outcome met the real requirement.
Watch on YouTube
