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

## Executive summary

The presentation details the evolution of LLM agents, demonstrating a shift from complex, brittle Python code loops to declarative, file-based definitions using system instructions and skills. The speaker shows that modern agent architectures, such as the Gemini API's anti-gravity agent, utilize a hosted sandbox and network proxy to manage state and credentials securely. This allows agents to operate using general-purpose tools (like GitHub CLI or Google Search) defined in files (e.g., `AGENTS.md`), drastically reducing the need for thousands of lines of custom orchestration code.

## Key takeaways

- The Agent Evolution: Code to Files: Agent development is moving away from writing explicit Python loops, JSON schemas, and tool routing logic. The core functionality is now expressed in files (Markdown/Skills) that define instructions, rules, and capabilities, allowing the model to use general tools.
- Server-Side State Management: The new architecture handles complex tasks by moving loops, tool routing, session state, and context compaction to the server side, requiring only a single API call with new inputs.
- Security and Isolation: A hosted sandbox and network proxy ensure that the agent never sees the actual credentials, injecting tokens only when outbound requests are made, and allowing domain restriction for enhanced security.
- Focus on Domain Logic: The primary work for developers is now defining the domain instructions, rules, and evaluation criteria (the 'what'), rather than writing the infrastructure code (the 'how').

## Technical details

- Interactions API (Gemini API): This unified interface supports running models and agents, featuring server-side state management and background execution. It replaced the traditional turn-based conversation history with a 'steps' timeline (user input, reasoning, function call, function result).
- Agent Architecture Shift: The process moved from requiring explicit Python loops, defining JSON schemas, and writing tool implementations to using files (e.g., `AGENTS.md`) to provide system instructions and capabilities (Skills).
- Sandbox Environment: The anti-gravity agent runs within a hosted, isolated cloud sandbox, enabling it to execute bash commands and save files, while the network proxy secures credentials by injecting tokens on outbound requests.
- Agent Definition: Agents can be defined using a custom ID, system instruction, base agent, and environment, allowing existing code and configurations to be reused while leveraging the new API structure.

## Practical implications

- Build engineers can significantly reduce the amount of boilerplate orchestration code required for complex agent workflows.
- The focus shifts from managing execution paths (loops, state) to defining clear, declarative rules and capabilities in configuration files.
- The use of a secure, isolated sandbox and network proxy simplifies CI/CD integration by abstracting away direct credential handling.
- The ability to extend agents by simply providing more files (Skills) rather than rewriting code streamlines feature development and maintenance.

## Topics

AI Agents, LLM Orchestration, Build Engineering, API Design, Serverless Computing, Interactions API, Gemini API, GitHub CLI, AGENTS.md

Source: https://www.youtube.com/watch?v=fjF8EKnxKCU