Hello community! This may have been discussed before but asking anyway, I am building out some voice agent flows.
Has anyone created a feedback loop that could train the agent based on assessing transcript and then update the agent call prompt with the learnings?
In my mind it is: agent calls -> based on transcript update a sheet row with assessment of what could be improved for smoother customer experience -> when sheet has a new row updated, update the prompt in the AI calling agent cc Marvin A. / Alon J.
you could also do this within the same agent btw. record meeting → analyze transcript → set node config
In terms of self updating prompts - phone agents are pretty particular and I wouldn't recommend too much prompt adjustment unless it was a very sound/minimal self healing loop - the tools Marvin shared would allow you to do this 🔥 Some easier alternatives:
Use the get task details to report any unpositive calls, take those learnings and manually filter + edit prompt!
Strongly recommend trying out the get task details action it is awesome. Has the full block context, input outputs, and more. Extremely useful for evals and debugging
Adding new knowledge as a memory / KB / google doc when a unique question shows up so things only have to be repeated once. If this is ever used load it all in PRIOR to the phone call step to save latency!
we typically see this for customer support type use cases
