Does hiring an agent beat doing the job yourself?

4 tasks, each done both ways. Only policy= differs between the columns — same driver, same tape, same cost model, held fixed by construction.

COUNTERFACTUAL — neither position was heldchain tape

Across all 4 tasks

no verdict (4 observations, need 30)

4 tapes are not evidence about a strategy. Each task’s own quote is, and it is drawn from many windows.

tasks: 4
quotable: 4
withheld: 0
agent ahead: 0
DIY ahead: 2
too close to call: 2
bands separated: 2

The 4 tasks

The baseline and the tape both differ per task. 3 on 31.0 days, 1 on 17.0 days. Each column is compared against the other on that task’s own tape, so the window bounds how much of a year the annualised figure describes — not which of the two is ahead.

trading · PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050

Earn — fees on a liquidity position

−64.29pp

Without an agent: mint once at the same width, never touch it (passive_policy)
With an agent: Warden — Avellaneda–Stoikov recentring

Warden — Avellaneda–Stoikov recentring-52.10% – -49.97%

mint once at the same width, never touch it (passive_policy)4.78% – 27.21%

The bands are separated — the gap survives the spread, not just the medians. −64.29pp at the median.

−64.29pp agent loses to DIY by 64.29pp, bands do not overlap

fees earned: 0.00994223 vs 0.0175265 — worse

material · bands separated · 31.0 days of tape

The supporting numbers

Earn — fees on a liquidity position: agent against baseline
Metric	Warden — Avellaneda–Stoikov recentring	mint once at the same width, never touch it (passive_policy)
median return	-51.24%	13.05%
P25 – P75	-52.10% – -49.97%	4.78% – 27.21%
in range	6.1%	6.6%
fees (WBNB)	0.0099	0.018
adverse selection (upper bound) (WBNB)	0.0026	0.018
costs (WBNB)	0.047	0.00078
moves	498	1

net return on capital (fees − realized convexity cost − costs), P25–P75 · on 1.00 WBNB of capital · annualised

security · PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050

Protect — avoid being picked off by one-way flow

−38.17pp

Without an agent: the same wide band, held through everything (withdrawal disabled)
With an agent: Sentinel — withdraw on §3.4 toxicity, re-enter after m_clear

Sentinel — withdraw on §3.4 toxicity, re-enter after m_clear-24.16% – -23.85%

the same wide band, held through everything (withdrawal disabled)9.38% – 14.94%

The bands are separated — the gap survives the spread, not just the medians. −38.17pp at the median.

−38.17pp agent loses to DIY by 38.17pp, bands do not overlap

realized convexity cost (upper bound on LVR): 0.000464581 vs 0.00854537 — 18.4x better

material · bands separated · 31.0 days of tape

The supporting numbers

Protect — avoid being picked off by one-way flow: agent against baseline
Metric	Sentinel — withdraw on §3.4 toxicity, re-enter after m_clear	the same wide band, held through everything (withdrawal disabled)
median return	-24.07%	14.10%
P25 – P75	-24.16% – -23.85%	9.38% – 14.94%
in range	9.5%	100.0%
fees (WBNB)	0.0017	0.019
adverse selection (upper bound) (WBNB)	0.00046	0.0085
costs (WBNB)	0.020	0.0015
moves	498	2

net return on capital, P25–P75 (the cost of the withdrawals is charged in full) · on 1.00 WBNB of capital · annualised

security · two venues: PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050 vs PancakeSwap v3 WBNB/USDT 0.25% · 0x1401ff943D08a7E098328C1d3a9d388923B115D2 — depth and the flow screen chose the same one

Choose — which pool to provide liquidity to

±0.00pp

Without an agent: pick the deepest pool — PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050, 84.0x the median liquidity (the obvious heuristic: more TVL is safer)
With an agent: pick the pool whose flow is not one-way — PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050, where §3.4's imbalance arm fires on 41.9% of samples against 42.8% on the other — which is the pool depth chose too

pick the pool whose flow is not one-way — PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050, where §3.4's imbalance arm fires on 41.9% of samples against 42.8% on the other — which is the pool depth chose too-620.55% – -617.83%

pick the deepest pool — PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050, 84.0x the median liquidity (the obvious heuristic: more TVL is safer)-620.55% – -617.83%

The bands overlap — at this sample size the two are not distinguishable, whatever the gap between their medians.

±0.00pp indistinguishable: +0.00pp is below the 0.10pp materiality floor

immaterial · bands overlapping · 31.0 days of tape

The supporting numbers

Choose — which pool to provide liquidity to: agent against baseline
Metric	pick the pool whose flow is not one-way — PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050, where §3.4's imbalance arm fires on 41.9% of samples against 42.8% on the other — which is the pool depth chose too	pick the deepest pool — PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050, 84.0x the median liquidity (the obvious heuristic: more TVL is safer)
median return	-619.71%	-619.71%
P25 – P75	-620.55% – -617.83%	-620.55% – -617.83%
in range	6.1%	6.1%
fees (WBNB)	0.00032	0.00032
adverse selection (upper bound) (WBNB)	0.000084	0.000084
costs (WBNB)	0.016	0.016
moves	498	498

net return on capital of the same agent on the chosen venue, P25–P75. Capital for this task is 0.03181, not the report's 1. A1's ceiling is eps x pool liquidity and therefore a property of the pool, and the shallower venue binds it — PancakeSwap v3 WBNB/USDT 0.05% · 0x36696169C63e42cd08ce11f5deeBbCeBae652050: 7.031 · PancakeSwap v3 WBNB/USDT 0.25% · 0x1401ff943D08a7E098328C1d3a9d388923B115D2: 0.06362 (at 0.5x margin). Both columns run at the same figure, because a comparison between two venues at two capitals is partly a comparison between two position sizes. · on 0.032 WBNB of capital · annualised

trading · Venus Core Pool (BSC) — vUSDT and vUSDC, verified three ways

Route — which lending venue to supply to

−0.01pp

Without an agent: supply to the highest-rate venue once and never move (park_policy)
With an agent: Router - move only when the rate edge clears the round-trip cost

Router - move only when the rate edge clears the round-trip cost1.71% – 1.92%

supply to the highest-rate venue once and never move (park_policy)1.90% – 1.92%

The bands overlap — at this sample size the two are not distinguishable, whatever the gap between their medians.

−0.01pp indistinguishable: -0.01pp is below the 0.10pp materiality floor

moves made (best rate seen 2.98%, hurdle 1.06%, break-even at 2 days): 2 vs 1 — worse

immaterial · bands overlapping · 17.0 days of tape

The supporting numbers

Route — which lending venue to supply to: agent against baseline
Metric	Router - move only when the rate edge clears the round-trip cost	supply to the highest-rate venue once and never move (park_policy)
median return	1.90%	1.91%
P25 – P75	1.71% – 1.92%	1.90% – 1.92%
in range	0.0%	0.0%
fees (WBNB)	0.00	0.00
adverse selection (upper bound) (WBNB)	0.00	0.00
costs (WBNB)	1.03	0.015
moves	2	1

net return on supplied capital (realized yield - switch costs), P25-P75, over a 168h stated holding period. Capital for this task is $10,000 of stablecoin, not the report's figure — that one is denominated in WBNB and means nothing to a lending market. · on 10,000.00 WBNB of capital · net return on supplied capital (realized yield − switch costs)

The same report, on too little history

The same code path on a tape too short to reach the policy horizon. It produces no numbers at all, which is the correct answer.

Deliberately short tapesynthetic tape3 of 3 withheld

Earn — fees on a liquidity position
no verdict — baseline: 0 usable replays, assumption A5 requires 20; 60 window(s) were shorter than the 24h policy horizon; agent: 0 usable replays, assumption A5 requires 20; 60 window(s) were shorter than the 24h policy horizon
Protect — avoid being picked off by one-way flow
no verdict — baseline: 0 usable replays, assumption A5 requires 20; 60 window(s) were shorter than the 24h policy horizon; agent: 0 usable replays, assumption A5 requires 20; 60 window(s) were shorter than the 24h policy horizon
Choose — which pool to provide liquidity to
no verdict — baseline: 0 usable replays, assumption A5 requires 20; 60 window(s) were shorter than the 24h policy horizon; agent: 0 usable replays, assumption A5 requires 20; 60 window(s) were shorter than the 24h policy horizon

Overall: no verdict (0 observations, need 30)

How each of these quotes was constructed →