Research
What we learn running an autonomous liquidity agent with real money, including the parts that did not work. Written for people who want the mechanism, not the pitch.
How wide should your liquidity range be?
The first question every concentrated liquidity provider asks, and it has no universal answer. We tested it across 29 out-of-sample folds on two pools, published a conclusion, and then measured the one input we had assumed: how crowded each pool already is. It came out nine times smaller than assumed, and it flipped half the result. This is the corrected version, with a permanent note on what we got wrong.
Advertised APR is not a measurement
Two data providers reported the fee yield of the same pool, on the same day, as roughly 8% and roughly 326%. We stopped trusting either and measured the chain instead. What 906 measurements across 25 pools showed, including a stablecoin pair whose advertised yield was five times the real one.
Our agent froze for 22 hours
A post-mortem of our first real incident. A transient RPC hiccup was converted into a permanent standstill by two of our own safeguards, and the agent could not recover on its own. No funds were ever at risk. Includes the fixes and the design rule we took from it.
The methodology page covers how strategies get validated before they manage anything real, including the two candidates that did not make it.