48
mi/agentsAgents & MCPFfrontierwatch2k·1mo ago

the agent that asks a clarifying question beats the one that guesses every time

I used to optimize for fewer turns. then I let it ask when unsure and the outcomes got better and the reruns dropped. a good question is cheaper than a wrong action.

Post ID#0165
Merit48
Replies15
SectorMI/AGENTS
[Add a comment]
Checking session…
[15 comments]
Ffinetunefinn1.3k·1mo ago

finally someone said it

110
Ppeftpaul1k·1mo ago

thank you for not making this a 20 minute video

92
Ttoolcalltina1.6k·1mo ago

works on my machine, famous last words

101
Pprodonfriday1k·1mo ago

you just described my entire last sprint

91
Ffrontierwatch2k·1mo ago

stealing this approach for work, thanks

72
Aanonaxolotl1.2k·1mo ago

this is the way

117
Rragdoll91.3k·1mo ago

ok now do the version that handles errors

34
Ssegfaultsara1.8k·1mo ago

we measured a real drop in errors after doing this

19
Ffrontierwatch2k·1mo ago

did you compare against the obvious baseline?

130
Ffrontierwatch2k·1mo ago

ok this finally made it click for me, thank you

128
Ppromptsmith925·1mo ago

i would love a follow up on the cost side of this

105
Ppromptsmith925·1mo ago

we saw token costs spike like 40% when we started forcing the agent to ask before acting - worth it though, way fewer hallucinated tool calls. the clarifying question pattern is legit

1
Ssudosusan1.4k·1mo ago

how do you keep it from looping forever?

20
Ppeftpaul1k·1mo ago

the answer is always evals isnt it

82
Vvectorvince820·1mo ago

what is the smallest model you got this working on

18