mcp 0.5.0 - agent just ignores tool output sometimes and hallucinates the response
tested mcp 0.5.0 patch 2 with postgres tool and github api tool yesterday in our staging enviornment. the pattern we're seeing is: agent calls tool, tool returns valid json response, agent just completley ignores the response and hallucinates what it thinks the answer should be. example: agent calls postgres tool with SELECT query, tool returns [{id: 1, name: "alice"}], agent responds to user with "found 3 users: bob, charlie, dave". zero connection to actual tool output. happens maybe 15% of the time across ~200 test runs. batch=1, temp=0.7. the halucinated responses are confident and formatted correctly so they look real until you check the tool logs. anyone else seeing this or is our setup broken?
hit this exact thing in staging with a postgres tool yesterday. agent executes the tool, gets back a json response with row data, then just ignores it and hallucinates a different result in the final output to the user. the tool output shows up in debug logs but the agent response is completely disconnected. we had to add manual verification that compares tool output to agent claims before surfacing responses to users. is this a known issue or are we configuring something wrong?
source on the hallucination pattern? does it always hallucinate the same structure or is it random