Agent Experts: Finally, Agents That ACTUALLY Learn

Too Long; Didn't Watch — Summary

This video addresses the critical flaw in current AI agents—their inability to learn and retain information—by introducing "agent experts" that automatically acquire and reuse expertise, leveraging meta-agentics (meta prompts, agents, and skills) to build self-improving systems demonstrated through codebase management and multi-agent orchestration.

Main Takeaways

• Current AI agents suffer from a fundamental problem: they execute and forget, unlike traditional software that learns and improves with use.
• Agent experts are designed to overcome this by automatically learning from actions, storing expertise in evolving "mental models" (expertise files), and reusing this knowledge without human intervention.
• Meta-agentics, comprising meta prompts, meta agents, and meta skills, are foundational tools that enable the creation of systems that build and improve themselves.
• The video demonstrates agent experts (e.g., database and WebSocket experts) operating within a multi-agent orchestration system, validating their understanding against actual code and self-improving.
• A key three-step workflow for agent experts involves planning, building, and self-improving, where the self-improvement phase updates the agent's expertise files.
• Scaling compute by deploying multiple agent experts against a single problem enhances the confidence and accuracy of the solutions.

Detailed Summary

The Problem with Current Agents (0:00 - 1:32)

• Traditional software improves by storing user analytics, usage data, and patterns to create better algorithms.
• A massive problem with current AI agents is that they forget and, consequently, do not learn.
• Existing solutions like memory files, prime prompts, sub-agents, and skills require manual updates, consuming significant human time and effort.
• Memory files are global forced context that always loads, and expertise often requires breaking rules, which static memory files hinder.
• The goal is to teach agents to act, learn, and reuse expertise at runtime, creating "agent experts."

Introducing Agent Experts (1:33 - 3:13)

• The core difference between a generic agent and an agent expert is that one executes and forgets, while the other executes and learns.
• Agent experts automatically turn actions into expertise, storing relevant information and reusing it without human intervention.
• True experts continuously learn and update their "mental model," which is a data structure that evolves with each useful action.
• Experts accumulate information, examples, and expertise around specific topics, focusing on problems most relevant to the user, business, and customers.
• An expert's learning process never ends unless they stop learning, emphasizing the continuous improvement aspect.
• An agent expert is defined as a concrete form of a self-improving template metaprompt.

Meta Agentics: The Foundation (3:14 - 6:49)

• Meta-agentics involve building elements of the system that construct the system itself, including meta prompts, meta agents, and meta skills.
• These tools significantly increase an agentic engineer's output, whether or not agent experts are used.
• A meta prompt can create new versions of prompts, such as a "question with mermaid diagrams" prompt.
• A meta agent can build new agents, like a planner agent that reads and executes a plan prompt, enabling parallel planning.
• A meta skill can create new skills, such as a "start orchestrator skill" to kick off multi-agent orchestration applications.
• Meta-agentics are crucial for any codebase, serving as a foundation for quickly spinning up new agentic layers.
• The created "question with mermaid diagrams" prompt includes a structure for user questions and a purpose to always create mermaid diagrams.
• The meta agent built a planner agent that reuses existing prompts as sub-agents, scaling their utility.
• The meta skill converted an existing process into a concrete skill for starting the orchestrator front and back ends.

The Learning Component (6:50 - 7:27)

• While meta-agentics (meta prompts, meta agents, meta skills) enable action, they are not agent experts because they do not learn automatically.
• No automatic updates or proficiency improvements occur within these meta tools.
• A key characteristic of an agent expert is its ability to learn on its own, accumulating and managing its own expertise after initial instruction.

Agent Experts in Action: Database Expert (7:28 - 11:52)

• A concrete example of an agent expert is a database expert, which has its own expertise file.
• This expertise file serves as the agent expert's mental model of its problem space.
• When asked a question (e.g., "How does information flow between our database tables?"), the agent first reads its expertise file.
• It then instantly compares and validates its understanding against the actual codebase, which is the true source of truth.
• The expertise file is a working memory file, not a source of truth, similar to a human engineer's mental model.
• The agent generates a comprehensive report, including an entity-relationship diagram and information flow patterns, for a multi-agent orchestration application.
• The agent's question prompt includes a static variable referencing the expertise file, ensuring it's read first.
• The workflow involves reading the expertise file, validating assumptions against the codebase, and then reporting.
• A "self-improve database" command allows the agent to synchronize its mental model with the working codebase.

Agent Experts in Action: WebSocket Expert (11:53 - 18:50)

• A WebSocket expert is demonstrated using a three-step agentic workflow: plan, build, and self-improve.
• The agent is tasked with adding a session-based counter to the app navbar for WebSocket events, showcasing its ability to update code.
• Multiple WebSocket experts can be deployed (e.g., three agents) against a single question to co-locate on the right answer, increasing confidence in the results.
• The orchestrator agent designs specialized system prompts for these experts.
• The three-step workflow involves a planning step, a build step against the plan, and a self-improve step.
• The self-improve step is critical for the agent to keep track of changes and update its expertise, preventing the engineer from manually tracking updates.
• The self-improve prompt for the WebSocket expert has a similar structure, suggesting it might have been generated by a meta-expert.
• The application's UI updates to show the WebSocket event counter, demonstrating the agent's successful code modification.
• The orchestrator synthesizes results from multiple agents, noting instances where some agents might not complete the task, highlighting the value of scaling compute.
• The plan step can consume a large number of tokens (e.g., 80k tokens), but the top-level agent's context is protected by delegating work to sub-agents.
• The build step can also be token-intensive (e.g., 61k tokens, 41 tool uses).
• The self-improve WebSocket expert kicks off, making updates as things change in the system.
• The foundational units for building these systems are context, model, prompt, and tools (the core four).

Notable Quotes

"The massive problem with agents is this. Your agents forget. And that means your agents don't learn." "The difference between a generic agent and an agent expert is simple. One executes and forgets, the other executes and learns." "Real experts don't relearn their craft every time they have a new task. True experts are always learning. They're updating their mental model." "The expert understands that the game they're operating in never ends except for one condition. The moment they stop learning." "The agent expert is a concrete form of a self-improving template metaprompt." "At their core, meta-agentics help you build elements of the system that builds the system." "Agent experts must learn on their own." "The expertise file is the mental model of the problem space for your agent expert." "The true source of truth is always the code. You and I know that. Not the comments, not the documentation, not the plans. It is the code." "Three times marks a pattern." "Context, model, prompt, tools, the core four."