How I would define an AI agent at the start of 2025:

In abstract, a piece of software that can achieve a certain goal, and uses AI to navigate the fuzzy path of achieving that goal.

More specifically, a combination of
- Goal
- Context management
- Underlying model
- Tools (datastores and functions with effects)
- Prompting & self-prompting strategies (including delegating to other agents)

Example: A geocoding agent

Goal: Parse content from a sentence and geocode it
Context Management: Trimmed sliding window. We only need the very most recent messages to get the job done.
Underlying model: A simple model like 4o mini is more than enough
Tools: A function to search cities with inbuilt query match ranking, and a  function with access to a Google Maps geocoding API
Prompting: Single turn. The agent should be able to get the job done in a single exchange, or ask the user to clarify the location if needed

Die klare Strukturierung eines KI-Agenten mit Ziel, Kontext, Modell und Werkzeugen zeigt, wie pragmatisch und effektiv KI heute eingesetzt werden kann. Besonders spannend finde ich die Rolle des Promptings für flexible Interaktionen.