Hey team, incredible work on the mcp-agent framework! We've been looking deeply into native MCP integration patterns.
One major bottleneck we've seen scaling standard MCP agents is Context Drift—when the external state changes but the agent's context window holds stale representations, leading to hallucinated tool calls.
We recently solved this on our end (Smriti) by building a stateless proxy that intercepts ools/call requests and injects a _temporal_metadata signature (execution timestamp + run/task IDs) into the payload. This allows the agent to guarantee point-in-time state recovery without adding background daemons.
Would you be open to a PR or discussion on standardizing a Temporal Signature for mcp-agent to make stateful workflows deterministic?
Hey team, incredible work on the mcp-agent framework! We've been looking deeply into native MCP integration patterns.
One major bottleneck we've seen scaling standard MCP agents is Context Drift—when the external state changes but the agent's context window holds stale representations, leading to hallucinated tool calls.
We recently solved this on our end (Smriti) by building a stateless proxy that intercepts ools/call requests and injects a _temporal_metadata signature (execution timestamp + run/task IDs) into the payload. This allows the agent to guarantee point-in-time state recovery without adding background daemons.
Would you be open to a PR or discussion on standardizing a Temporal Signature for mcp-agent to make stateful workflows deterministic?