agent burned $47 in api calls retrying the same broken mcp tool call 83 times

working on an agent that uses 6 mcp tools (file ops, web search, postgres queries). last night it got stuck in a retry loop on a file write tool that was timing out due to permissions error on the server. agent retried with IDENTICAL args 83 times over 4 hours before i noticed and killed it. the error message from the server was clear ("permission denied writing to /var/log/app.log") but the agent just kept retrying. we don't have any client-side dedup logic right now. questions: 1. should we add a hash check on the client (tool name + args) and block retries within some window? 2. or is this a prompt engineering problem - should the system prompt explicitly say "don't retry with identical args after a failure" 3. what window makes sense for blocking retries - 60 seconds? last 5 tool calls? curious how others handle this