# Agent Experts: WHAT IF Your Claude Code Agents could LEARN? Source: https://www.youtube.com/watch?v=zTcDwqopvKE Recap page: https://rapidrecap.app/video/zTcDwqopvKE Generated: 2025-12-15T14:34:04.008+00:00 --- ## Quick Overview Agent Experts must learn on their own by using a self-improving workflow that chains planning, building, and self-improving steps, which is necessary because relying solely on static expertise files or an orchestrator agent results in agents that forget context and fail to scale effectively. **Key Points:** - Agents of today have many problems, most of which can be solved with great context engineering or agentic prompt engineering, but one massive problem persists: agents forget. - Traditional software improves as it's used by storing analytics and usage data to create algorithms, but current agents lack this continuous learning capability. - The core solution involves teaching agents to 'Act' and 'Learn' so they can 'Reuse' their expertise at runtime, rather than starting fresh every time a new task arises. - Agent Experts are defined by a 'Self-Improving Template Meta Prompt' that allows them to automatically update their own context, related prompts, or isolated files with new information, achieving expertise without human intervention in the loop. - The self-improvement process involves a four-step workflow: Plan, Build, Self-Improve, and Report, ensuring that the agent's mental model is constantly validated against the codebase (the true source of truth). - The WebSocket Expert example demonstrates this by running three agents concurrently to answer a question, utilizing the self-improvement mechanism to validate and update its expertise based on the code it reads. - The ultimate goal of Agent Experts is to move beyond static context/expertise files and create systems where agents continually accumulate and refine their knowledge base, leading to exponentially increasing performance over time. ![Screenshot at 00:05: The video introduces the concept of Agentic Prompt Engineering as the solution to problems faced by today's agents, highlighting the need to move beyond simple context engineering.](https://ss.rapidrecap.app/screens/zTcDwqopvKE/00-00-05.png) **Context:** This video introduces the concept of 'Agent Experts' as an evolution beyond standard AI agents, focusing on overcoming the critical problem of 'forgetting' context between sessions. The presenter contrasts this with traditional software, where usage data helps algorithms improve over time, arguing that current agents need a similar mechanism to build lasting expertise rather than resetting with every interaction. ## Detailed Analysis The video argues that the biggest problem with current AI agents is their inability to retain and build upon knowledge over time, unlike traditional software that uses user analytics to improve its algorithms. The solution proposed is 'Agent Experts,' which utilize a 'Self-Improving Template Meta Prompt' that orchestrates a four-step workflow: Plan, Build, Self-Improve, and Report. This process allows agents to learn from their actions (Act/Learn/Reuse cycle) and continuously update their internal 'mental model' by comparing their existing expertise against the actual codebase (the 'true source of truth'). The expert agents use this self-validation loop to ensure their knowledge remains accurate, which is demonstrated using a 'WebSocket Expert' that scales up to run three specialized agents in parallel to answer a complex question about system events. The key takeaway is that Agent Experts encode their expertise directly, allowing them to perform better than generic agents that start fresh every time, enabling scalable and continuously improving AI systems. ### Agent Problems - Agents of today have many problems - Most of them can be solved with great context engineering or great agentic prompt engineering - One massive problem that persists is that your agents forget ### The Learning Gap - Traditional software improves as it's used by storing user analytics and patterns that create algorithms - Your agents don't learn ### The Agent Expert Solution - Agents must learn on their own - You teach them to Act, Learn, and Reuse their expertise at runtime - This is achieved via a Self-Improving Template Meta Prompt ### The Self-Improvement Workflow - Four steps: Step 1: Create Plan - Step 2: Build - Step 3: Self-Improve - Step 4: Report - Each step uses a subagent to perform specific tasks, ensuring validation against the codebase. ### Demonstration (WebSocket Expert) - The WebSocket Expert uses a self-improvement workflow to analyze existing expertise against the codebase, resulting in a detailed report on WebSocket events and outperforming agents that rely only on static context. ![Screenshot at 00:05: The video introduces the concept of Agentic Prompt Engineering as the solution to problems faced by today's agents, highlighting the need to move beyond simple context engineering.](https://ss.rapidrecap.app/screens/zTcDwqopvKE/00-00-05.png) ![Screenshot at 00:24: The core challenge is introduced: agents forget, contrasting with traditional software that learns from usage data.](https://ss.rapidrecap.app/screens/zTcDwqopvKE/00-00-24.png) ![Screenshot at 01:22: The proposed solution involves a cycle of Act, Learn, and Reuse, allowing agents to build expertise over time rather than starting fresh.](https://ss.rapidrecap.app/screens/zTcDwqopvKE/00-01-22.png) ![Screenshot at 02:23: The structure of an 'Expert's Mental Model' is visualized, showing layers of expertise built around a core 'Topic Focus' using context, information, examples, and patterns.](https://ss.rapidrecap.app/screens/zTcDwqopvKE/00-02-23.png) ![Screenshot at 03:11: The definition of the 'Self-Improving Template Meta Prompt' is shown, emphasizing its ability to automatically update related files and prompts based on new information.](https://ss.rapidrecap.app/screens/zTcDwqopvKE/00-03-11.png)