MCP-Cosmos: World Model-Augmented Agents for Complex Task Execution in MCP Environments

MCP-Cosmos is a framework that integrates generative World Models (WM) into the Model Context Protocol (MCP) ecosystem to enable predictive task automation.

It uses a 'Bring Your Own World Model' (BYOWM) strategy, allowing agents to simulate state transitions and refine plans in a latent space before execution.

The framework was evaluated on 24 MCP-Bench tasks (300+ trajectories) using two agent strategies (ReAct and SPIRAL) with two planning models (GPT-OSS-120B, CLAUDE-SONNET-4.6) and three world models (GPT-OSS-120B, CLAUDE-SONNET-4.6, Arctic-AWM-4B).

Results show that world-model-augmented agents (ReAct-Plan-Exec and SPIRAL-Exec) outperform the ReAct baseline in tool selection, parameter accuracy, and parallel efficiency, with SPIRAL-Exec + GPT-OSS-120B-WM achieving the highest overall score (44.8% vs.

36.1% baseline).

However, the baseline ReAct had higher task fulfillment and dependency awareness.

The paper identifies a gap in current evaluation metrics and proposes a new metric, Execution Quality, which combines tool call success rate and normalized average tool calls to penalize excessive retries.

Using this metric, world-model agents score significantly higher (e.g., 91.4% for SPIRAL-Exec + GPT-OSS-120B-WM vs.

38.9% for ReAct).

An ablation study shows that a stronger planner (CLAUDE-SONNET-4.6) without a world model issues 29.78 tool calls per task (vs.

7.04 for GPT-OSS-120B), but world models constrain this exploration, improving efficiency.

Limitations include static environments, computational overhead, and non-portable metrics.