Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind
Philipp Schmid from Google DeepMind shows how to build agents — AI systems that complete tasks autonomously — with progressively less code.
Version one uses a hand-written Python loop with error handling. Version two uses a framework that auto-generates schemas. Version three has no Python at all: the agent runs in a secure sandbox on the server and accesses tools without seeing passwords or tokens. Instead of code, you write instructions in text files. Schmid's core message is that if your agent code grows more complex as models improve, you're overengineering — teams have discarded thousands of lines of orchestration code in favor of just a few hundred lines of markdown.
Vibekollen prepared this summary with AI from the original publication. The content belongs to AI Engineer.
