Alita: Generalist Agent Enabling Scalable Agentic Reasoning with Minimal Predefinition and Maximal Self-Evolution

Alita achieves 75.15% pass@1 and 87.27% pass@3 accuracy, which is top-ranking among general-purpose agents, on the GAIA benchmark validation dataset.

The reliance on large-scale manually predefined tools and workflows introduces several critical limitations:

1. Incomplete Coverage: It is impractical, if not impossible, to predefine all the tools required for the wide variety of real-world tasks an agent might encounter.
2. Limited Creativity and Flexibility: Many complex tasks require agents to creatively compose new tools or leverage existing ones in novel ways, while pre-designed workflows and hardcoded components constrain this compositional flexibility and inhibit the development of adaptive behaviors.
3. Mismatch: It is not always the case that the interface or environment of different tools is compatible with the agent. For example, many useful tools are not written in Python, which makes it difficult, though not entirely impossible, for them to be pre-connected to the mainstream agent frameworks that are primarily written in Python.

Together, these challenges ultimately hinder the scalability, adaptability, and generalization of existing generalist agents.

In contrast to the prevailing trend of growing complexity, we propose a radically simple design philosophy built on two principles:

1. Minimal Predefinition: Equip the agent with only a minimal set of core capabilities, avoiding manually engineered components for specific tasks or modalities.
2. Maximal Self-Evolution: Empower the agent to autonomously create, refine, and reuse external capabilities as needed.

We instantiate this vision through Alita, a generalist agent built with a single core capability (i.e., the web agent) and a small set of general-purpose modules that enable self-directed capability expansion. Specifically, we take advantage of the Model Context Protocols (MCPs), which is an open protocol that standardizes how different systems provide context to LLMs, and empower Alita to dynamically generate, adapt, and reuse MCPs based on the demands of each task rather than relying on static, predefined tools. This shift from manually designed capabilities to on-the-fly MCP construction unlocks a new path for building agents that are simple yet profoundly capable.

Alita-generated MCP Box has two benefits.

1. Agent Distillation: The reuse of auto-generated MCPs can be viewed as a way of distillation, which is much cheaper and easier than traditional distillation.
   Stronger Agent teaches Weaker Agent: These MCPs can be reused by other weaker agents and improve their performance since Alita, instead of human developers, designs a set of useful MCPs fit to GAIA by trial and error.
   Agent with Larger LLMs teaches Agent with Smaller LLMs: These MCPs can also be reused by agents with smaller LLMs and significantly improve the performance.
2. Make Pass@1 approach Pass@N: The MCP Box can also be connected to Alita, and makes pass@1 approach Pass@N.