Enhance MCP Concurrency and Context Management

[021gink]

Description

As highlighted by Ronacher, MCP is not designed for "no-reasoning automation" and struggles with concurrency and context management. We propose the following enhancements based on his suggestions:
1.Concurrency Support: Enable MCP to handle multiple tasks simultaneously to improve efficiency.
2.Context Management Optimization: Reduce context window usage to allow more tasks to be processed.
3.Code Generation and Sandbox Execution: For no-reasoning tasks, provide an alternative code generation route that directly assigns tasks to local or cloud sandbox environments, similar to how Artifacts or Code Interpreters work. This approach can bypass the limitations of context window and enable true concurrency.

References

No response

[Kludex]

I'm a big fan of Armin, but this issue doesn't seem to propose anything concrete.

[021gink]

I agree with your perspective — the issue raised was rather poorly considered. Upon reflection, the context explosion problem is unavoidable, whether it's tool invocation or plain conversation. It's fundamentally constrained by the model's autoregressive nature.

As for the concurrency issue, MCP actually supports batch processing, so when designing the client, we can account for the need to repeatedly execute the same workflow concurrently. This isn't something the MCP protocol itself needs to address, especially since the one-to-one connection was originally designed for better isolation. In that sense, the concern is admittedly a bit silly.

Recently, there's been work on HRM that might drive a breakthrough in model architecture. It could potentially solve the compute explosion problem in video generation models. The autoregressive paradigm limits wider application of models — architectural evolution is needed.