Built a stateless A2A-to-MCP bridge with cross-instance task recovery

[aarushitandon0]

Hi everyone!

I've been working on [a2a-mcp-bridge](https://github.com/aarushitandon0/a2a-mcp-bridge), an open-source Python bridge that exposes A2A 1.0 agent skills as MCP tools.

The problem I wanted to explore was how to preserve access to long-running A2A tasks without maintaining task mappings inside the bridge.

For example, an A2A agent might pause a task to request additional input. If the bridge instance that initiated the request crashes, another instance should still be able to resume that same task without depending on the original process's memory.

I approached this using CBOR-encoded, AES-256-GCM-sealed task handles that carry the remote agent ID, task ID, context ID, and expiry. Any bridge instance with the appropriate configuration and encryption key can resolve the handle and continue interacting with the original A2A task.

Some of the implementation details:

• Agent Card discovery, with each advertised skill exposed as a namespaced MCP tool.
• Explicit mapping between A2A task states and MCP task statuses.
• Separate handling for INPUT_REQUIRED and AUTH_REQUIRED interruptions.
• Task resumption across bridge instances without a task database or shared cache.
• Support for rolling encryption key rotation.

I also tested the architecture with three bridge instances behind a round-robin load balancer. I terminated the instance that initiated a task, resumed it through another instance, and confirmed completion. The repository includes reproducible demonstrations for instance failure, full-pool restarts, and key rotation.

Repository: https://github.com/aarushitandon0/a2a-mcp-bridge

Reproducible demo: python demo/run_local.py (after installing the project)

I'd particularly appreciate feedback on the A2A task lifecycle integration, authentication handling, and compatibility considerations I might have missed when working with different A2A agents.

Thanks!