Are Agent Skills the New RAG?

Key Concepts

• Agent Skills: A modular, open-standard framework (popularized by Anthropic) that provides AI agents with procedural knowledge and specific workflows.
• Progressive Disclosure: A design pattern where only minimal metadata is loaded into the LLM's context window initially, with full instructions and files loaded only when a skill is triggered.
• Dynamic Context: The ability for an agent to discover and load capabilities on-demand rather than relying on a bloated, static system prompt.
• Code Execution Sandbox: An isolated environment (e.g., Docker containers) where an agent can safely execute arbitrary code to perform calculations or generate files.
• MCP (Model Context Protocol): A standard for connecting AI agents to external data sources, tools, and APIs.
• Daisy Chaining: The ability for an agent to trigger multiple skills sequentially within a single conversation.

---

1. The Evolution of Agentic Workflows

The video argues that while LLMs are powerful generalists, they lack the "procedural knowledge" required for specific company processes. Previous attempts to solve this—such as ad-hoc chatting, bloated system prompts, or static specialist agents—were inefficient.

• The Shift: The industry is moving from "prompt engineering" to "context engineering," where agents dynamically discover and execute skills as needed.
• The "Skill" Structure: A skill is essentially a folder containing a skill.md file (instructions/metadata) and optional sub-folders (scripts, references, assets).

2. Technical Implementation & Methodology

Progressive Disclosure Mechanism

To prevent context window bloat, the system uses a two-stage loading process:

1. Discovery Phase: The agent only sees the skill's name and a short description (YAML front matter).
2. Execution Phase: If the LLM determines a skill is required, it loads the full skill.md file and any necessary referenced files (e.g., Python scripts, branding guidelines).

Code Execution Sandboxes

To move from "chatting" to "doing," agents require an executable environment.

• Tooling: The author recommends LLM Sandbox, a lightweight, portable environment for running LLM-generated code.
• Security: Because Docker containers share the host kernel, the author suggests using gVisor as a security layer to prevent container escapes.
• Optimization: Pre-warming containers reduces latency, and custom images can be pre-configured with necessary libraries.

3. Common Skill Patterns

• Sequential Workflow Orchestration: Steering an LLM through multi-step processes (e.g., onboarding a customer: Create Account → Set Payment → Subscription → Welcome Email).
• Multi-Tool Coordination: Orchestrating various MCPs and APIs (e.g., Figma export → Drive upload → Linear task creation → Slack notification).
• Iterative Refinement: Using loops to validate outputs, perform quality checks, and refine drafts until they meet specific criteria.
• Context-Aware Tool Selection: Using business logic to decide where to store files or which API to call based on metadata like file size or user permissions.

4. Real-World Application: Customer Service Reporting

The author demonstrates a custom agent that generates monthly reports:

1. Trigger: The user asks for a "Customer Service Monthly Report."
2. Discovery: The agent identifies the relevant skill and loads the workflow.
3. Data Extraction: The agent uses a sub-agent to analyze weekly reports in an isolated context, preventing main-prompt pollution.
4. Calculation: The agent writes and executes Python code in a sandbox to aggregate metrics (ticket volumes, escalations).
5. Generation: The agent triggers a second skill to generate a formatted Word document using the calculated data.

5. Notable Quotes

• "A true specialist needs procedural knowledge. It needs to understand your company's workflows and your processes."
• "The single most important aspect of skills is the concept of progressive disclosure because you need to protect the context window of your main agent at all costs."
• "With RAG you were able to give your agent knowledge, whereas with skills you can provide the expertise... the step-by-step workflows and processes."

6. Synthesis and Conclusion

The integration of Agent Skills and Code Sandboxes represents a significant leap in AI utility. By moving away from static, bloated prompts toward a modular, open-standard approach, developers can create agents that are not only knowledgeable (via RAG) but also capable of executing complex, deterministic, and repeatable business processes. The ability to daisy-chain skills and delegate tasks to isolated sub-agents allows for a highly scalable architecture that remains performant within a single context window. For those looking to implement this, the author points to agentskills.io as the primary resource for the open standard.