ALITA-G self-evolves agents by generating and curating MCP tools

Researchers present ALITA-G, a self-evolution framework that transforms general-purpose agents into domain experts by systematically generating, abstracting, and curating Model Context Protocol (MCP) tools. The system synthesizes candidate tools from successful task trajectories, consolidates them into an MCP Box, and uses retrieval-augmented selection at inference time.

On the GAIA validation set, ALITA-G achieves 83.03% pass@1 and 89.09% pass@3, establishing a new state-of-the-art result.
The framework reduces mean tokens per example by approximately 15% relative to a strong baseline agent.
It demonstrates strong gains on benchmarks including GAIA, PathVQA, and Humanity's Last Exam.

ALITA-G provides a principled pathway from generalist capability to reusable, domain-specific competence, improving both accuracy and efficiency on complex reasoning tasks.

Benchmarks