From 36% to 100%: How Self-Improving Agents Write Their Own Skills — Rafal Wilinski, Runlayer
Summary
This talk outlines a framework for creating a self-improving enterprise by moving beyond single-agent intelligence. The core concept involves using 'skills'—reusable playbooks—to guide AI agents through complex, multi-step tasks. To scale this knowledge across an entire organization, the speaker proposes using MCP (Message/Command Protocol) as a unified distribution layer. By implementing a 'self-improving organizational flywheel,' successful and failed agent runs are distilled into skills, turning ephemeral, borrowed frontier intelligence into persistent, proprietary corporate knowledge.
Key takeaways
-
Skills are essential for deep work and reliability.
7:11
As agents become capable of working for hours, the initial trajectory is critical. Skills act as playbooks, guiding the agent and preventing it from wasting tokens or time by rediscovering known procedures, which is crucial for complex tasks (3:51, 4:31).
-
MCP provides the necessary centralized governance for skills.
12:35
To overcome the limitations of skills being local and client-specific, using MCP as a distribution layer allows the company to create a single source of truth for all skills, accessible to non-technical departments (7:55, 10:00).
-
The organizational flywheel converts failure into knowledge.
Knowledge should be distilled from both successful and failed agent runs. Failures are particularly valuable as they expose missing guardrails or edge cases, making the resulting skills more robust (11:59).
Technical details
-
Skills Definition and Function
212s
A skill is defined as a playbook that agents can read. It utilizes 'progressive disclosure,' meaning the agent first sees a summary, but can inject the full knowledge into the context when needed to change the task's trajectory (2:12).
-
Skill Distillation Process
The process involves taking a successful or failed agent run (a 'trace') and using a Large Language Model (LLM) to distill the sequence of tool calls into a reusable skill. This process is designed to be autonomous and asynchronous (11:34).
-
The Self-Improving Flywheel
The system groups similar agent runs, distills them into skills, and serves them via the MCP gateway. This creates a constantly evolving, centralized knowledge base that persists even if the underlying LLM models change or deprecate (15:23).
Mentioned resources
- Runlayer
- Voyager paper
- MCP (Message/Command Protocol)
Channel & topics
Watch on YouTube · Back to latest
This independent, AI-assisted summary is provided for commentary and informational purposes. It may contain errors or omit important context. Please watch the original video for the creator's complete presentation. Video, thumbnail, and related copyrights belong to their respective owners.