Paper 2512.16301
Adaptation of Agentic AI: A Survey of Post-Training, Memory, and Skills
- Published
- Dec 2025
- Research lab
- Independent
- Citations
- 15
- GitHub
- 679 stars
01 In brief
Summary
This survey examines how agentic AI systems are adapted after pretraining, organizing the field into a four-paradigm framework based on what is optimized (the agent or its tools) and the source of the adaptation signal (tool execution or agent output).
The paradigms are: A1 (tool-execution-signaled agent adaptation), A2 (agent-output-signaled agent adaptation), T1 (agent-agnostic tool adaptation), and T2 (agent-supervised tool adaptation).
The survey reviews post-training methods, adaptive memory architectures, and agent skills within this framework, comparing their trade-offs in cost, flexibility, and generalization.
It also summarizes evaluation practices across domains like deep research, software development, computer use, and drug discovery.
Key findings include that T2 tool adaptation can match A2 agent training accuracy with far fewer examples in retrieval settings, and that A1-trained agents can be frozen and redeployed as T1 tools.
The survey concludes by outlining open problems in agent-tool co-adaptation, continual learning, safety, and efficient deployment.
02 From the paper
Abstract
Large language model (LLM) agents are moving beyond prompting alone. ChatGPT marked the rise of general-purpose LLM assistants, DeepSeek showed that on-policy reinforcement learning with verifiable rewards can improve reasoning and tool use, and OpenClaw highlights a newer direction in which agents accumulate persistent memory and reusable skills. Yet the research landscape remains fragmented across post-training, retrieval, memory, and skill systems. This survey studies these developments under a single notion of \emph{adaptation}: improving an agent, its tools, or their interaction after pretraining. We organize the field with a four-paradigm framework spanning agent adaptation and tool adaptation. On the agent side, A1 (tool-execution-signaled) and A2 (agent-output-signaled) improve the agent itself through supervised fine-tuning, preference optimization, and reinforcement learning with verifiable rewards. On the tool side, T1 (agent-agnostic) provides reusable pre-trained modules any agent can call, while T2 (agent-supervised) uses the agent's outputs to train memory systems, skill libraries, or lightweight subagents. Using this framework, we review post-training methods, adaptive memory architectures, and agent skills; compare their trade-offs in cost, flexibility, and generalization; and summarize evaluation practices across deep research, software development, computer use, and drug discovery. We conclude by outlining open problems in agent-tool co-adaptation, continual learning, safety, and efficient deployment.