Static agents are dead. The future belongs to agents that adapt.
A new, comprehensive 65-page survey from top research institutions—including Stanford, Princeton, Harvard, and the University of Washington—has just provided the industry's first full taxonomy for Agentic AI Adaptation.
The paper argues that "Agentic AI" (large models that use memory, call tools, and act over multiple steps) is no longer just about execution. It is about evolution. The researchers map almost all advanced systems into just four basic patterns of adaptation.
The 4 Ways Agents Adapt (The Taxonomy)
The framework divides the world based on two questions: What gets updated? (The Agent or The Tools) and What is the signal? (Direct Results or Evaluations).
Type A1: The "Trial & Error" Agent
Mechanism: The agent is updated based on tool results.
Example: An agent writes code, runs it, sees an error, and updates its internal policy to avoid that syntax in the future. It learns from the direct reality of "Did this work?"
Type A2: The "Self-Correction" Agent
Mechanism: The agent is updated based on evaluations.
Example: An agent generates a plan, and a "Judge" model (or human) scores it. The agent updates its weights or prompt strategy based on that score. This is higher-level learning, focusing on quality rather than just execution.
Type T1: The "Better Library" Approach
Mechanism: The agent is frozen, but the retrievers are updated.
Example: You don't change the LLM; you change the RAG system. You train a retriever to fetch better documents or domain models. The "brain" stays the same, but the "textbook" it reads gets better.
Type T2: The "Sharpened Knife" Approach
Mechanism: The agent is frozen, but the tools are tuned from agent signals.
Example: The agent tries to use a search tool. Based on which search results actually helped the agent solve the task, the search tool itself is fine-tuned to surface those types of results faster next time.
The Trade-Offs
The paper explains that there is no "free lunch."
- A1 & A2 (Updating the Agent) offer maximum flexibility but come with high training costs and the risk of "catastrophic forgetting" (breaking old skills).
- T1 & T2 (Updating the Tools) offer modularity—you can swap out tools without retraining the brain—but they are limited by the frozen agent's inherent reasoning cap.
Sources
- Stanford, Princeton, Harvard, UW, et al. "Agentic AI Adaptation: A Survey." arXiv Preprint, 2025.
Disclaimer: This blog post reflects my personal views only. AI tools may have been used for brevity, structure, or research support. Please independently verify any information before relying on it. This content does not represent the views of my employer, Infotech.com.
Comments