Agentic Scholar

AI agents, tool-use, autonomous research systems, and the path to agentic intelligence.

MCP Protocol: Anthropic's Model Context Protocol for Tool-Using Agents

Anthropic · 2025–2026

An open standard connecting LLMs to external tools, data sources, and APIs. Standardizes how agents discover, authenticate, and invoke tools at runtime. Adopted by OpenAI, Google, and the broader ecosystem — becoming the "HTTP of AI agent tooling."

Code as World Models: LLMs Synthesizing Programs for Agent Planning

Multiple labs · 2025–2026

LLMs generate executable code to represent latent dynamics and observation models of a POMDP, then perform belief-state MCTS planning inside the synthesized world model. Significantly outperforms Chain-of-Thought agents on complex games and robotics tasks.

VAGEN: World Model Reasoning for Multi-Turn VLM Agents

NeurIPS 2025

Structures VLM agent reasoning into StateEstimation and TransitionModeling via RL in a POMDP formulation. A 3B model achieves 0.82 on 5 diverse agent tasks — nearly 3x improvement over untrained baseline, surpassing GPT-5 and Gemini 2.5 Pro.