Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind
A GitHub review agent is rebuilt with progressively less orchestration code, ending as instructions and setup files around general tools. Evals still remain the builder’s job.
The talk rebuilds one pull-request review agent three ways: a Python tool loop, an agent framework and a hosted filesystem agent. In the third version, the source directory disappears; the remaining agent definition is an **AGENTS.md** file plus a small GitHub CLI installer.
For new coding agents, start with general capabilities such as Bash, files and established CLIs before writing narrow tool wrappers. Move domain rules into skills and files, let the runtime manage loops and state, and keep **outcome evals** under your control.
The talk rebuilds one pull-request review agent three ways: a Python tool loop, an agent framework and a hosted filesystem agent. In the third version, the source directory disappears; the remaining agent definition is an **AGENTS.md** file plus a small GitHub CLI installer. For new coding agents, start with general capabilities such as Bash, files and established CLIs before writing narrow tool wrappers. Move domain rules into skills and files, let the runtime manage loops and state, and keep **outcome evals** under your control. The simplified version depends on a hosted sandbox, server-managed state, credential injection and capable models. Removing orchestration code does not remove the need for network restrictions, clean tools, verification or explicit behavioral rules.
This consolidates the filesystem-and-skill pattern into a design rule: begin with general tools and remove custom orchestration only when the runtime can safely supply state, isolation and credentials. It confirms the Hugging Face and Vercel examples, while narrowing their apparent simplicity by keeping evaluation, verification and security controls outside the reduced agent definition.