Agents Without Code: Skills, YAML, and Filesystems Replaced Python — Philipp Schmid, Google DeepMind

Philipp Schmid demonstrates the same GitHub pull-request reviewer in three forms: a custom Python tool loop, an agent framework, and a hosted agent configured with instructions and files. “Without code” means removing application-owned orchestration and custom tool wrappers, not eliminating the code in the hosted runtime or CLI tools.

This summary uses the video's English auto-generated captions. API behavior and security mechanisms below are Schmid's descriptions at the time of the talk, not independently verified current guarantees.

Main points

• A framework takes over the loop, but not every integration. The first implementation defines tool schemas, calls Python functions, handles errors, and feeds results back to the model. The framework version moves that repeated machinery into ADK and derives schemas from function signatures, but the developer still supplies tool implementations and their execution environment. Both demonstrated reviewers are limited to their configured tools; that is a property of these examples, not an inherent limit of Python or frameworks. 02:20–07:40(opens in a new tab)
• A hosted runtime shifts what the application owns. Schmid describes an Antigravity-based agent on the Gemini API with an isolated Linux sandbox, configurable source files, server-side session state, and automatic context compaction. In the third reviewer, instructions point the agent to the GitHub CLI and a small shell script installs it when needed. The runtime still executes the tool loop; the application supplies the task and environment rather than implementing that loop itself. 07:40–08:44(opens in a new tab), 09:48–14:57(opens in a new tab)
• General tools widen the available actions. Instead of calling dedicated functions for each GitHub operation, the hosted reviewer explores the sandbox and uses Bash, files, and the CLI. For weather, Schmid reports it using Google Search; he describes the ADK example as lacking a weather tool. This illustrates a broader tool set, not better PR-review accuracy. 06:36–07:40(opens in a new tab), 09:48–12:50(opens in a new tab)
• Credential handling and network scope remain design decisions. Schmid says a proxy injects credentials into outbound requests so the agent need not see the tokens directly, while domain configuration controls network access. His example configures GitHub credentials and also allows other web access without those credentials. He describes unrestricted network access as the default, so a hosted sandbox should not be read as a claim that every destination is blocked. 08:44–09:48(opens in a new tab), 11:47–13:54(opens in a new tab)
• Skills move task guidance into files. For a proposed security-scanning extension, Schmid contrasts adding a Python function and schema with supplying a skill file that explains which CLI to use, plus the tool in the environment if needed. He also describes writing preferences and task handoffs to files for later reuse. Files carry instructions and context; executable tools still perform the work. 13:54–15:59(opens in a new tab), 16:00–18:06(opens in a new tab)
• Keep domain knowledge and evaluations even as orchestration shrinks. His closing advice is to provide general tools, let the model choose execution paths, and focus on domain instructions, workflows, evaluations, and verified outcomes. He cites other teams' reductions in harness code as supporting anecdotes, not controlled evidence that deleting code always improves an agent. 14:57–18:06(opens in a new tab)