Report · measured 2026-09-09
Three tasks, each done twice
Once by asking an agent. Once by hand — every call a person would have to make, in the same minute, over the same node. The plan was to measure how much time an agent saves.
That is not what came back. The hand-done route answered 0 of 3 questions; the agent answered 3 of 3. On one task the hand-done path was quicker — because the public node refused its log queries, so it skipped the only expensive step and returned a number it already had. The fastest path in any comparison is always the broken one, which is why no ratio is published where a side failed to answer.
How the hand-done path was measured. It is a script issuing the same calls a person would have to issue — no page loads, no reading, no typing an address into a block explorer, no deciding which explorer to open. It is therefore a floor on what a person costs, never an estimate of it. Inflating that side would have been easy and would have made this document worthless.
What is recorded as a failure. A query the node refused is recorded as unmeasured, never as zero. “Nobody traded in this pool” and “we could not look” are different answers, and a report that renders them the same is worse than no report.
What can be re-run. Everything. node scripts/advantage-report.mjs writes the JSON this page is generated from; the endpoints it calls are public, need no key, and are the same ones any caller gets.
Task 1trading / liquidity provision
I have $10,000 to provide as liquidity for CAKE/WBNB on PancakeSwap. Which of the five fee tiers should it go into?
The pair lives in up to five pools at once. Every public interface ranks them by the capital already parked in them, and that is not what an LP is paid on.
Asking the agent
6,622 ms
1 request
Answered the question
Doing it by hand
968 ms
22 requests
Did not answer
the public node refused the log range, so no tier could be priced — only the capital ranking survived, which is the number every public interface already shows
No time ratio is published for this task. One of the two paths never produced the answer, and dividing a time by a non-answer would turn a failure into a benchmark.
What the five pools actually paid, over 59.3 minutes of chain
| Tier | Capital parked | Swaps | Fees paid | Fees per $1,000 parked |
|---|
| V2 0.25% |
$19,940,227.95 |
205 |
$194.92 |
$0.00978 |
| V3 0.01% |
$24,128.49 |
1623 |
$6.69 |
$0.27718 |
| V3 0.05% |
$1,062,742.35 |
929 |
$136.26 |
$0.12821 |
| V3 0.25% |
$5,278,552.66 |
82 |
$146.51 |
$0.02776 |
| V3 1.00% |
$3,516.53 |
0 |
$0 |
$0.00000 |
The tier holding the most capital is V2 0.25%. The tier that paid best is V3 0.01%. Every public interface ranks this pair by the first column; an LP is paid on the last one. Measured over a single window of 59.3 minutes and deliberately not annualised — forty minutes of flow says what happened in forty minutes.
Hired through the marketplace
#310460 Brain on BNB — PancakeSwap Fee Tier Placement — the same fee-tier measurement, delivered on-chain through the ERC-8183 escrow.
Asked through the hire path during this run, it quoted 0.1 $U (escrow ERC-8183, provider 0x7380…5963).
Task 2security / risk
Before I take a $2,500 position: what does that trade actually cost, and can the liquidity behind it be withdrawn tomorrow?
The headline slippage a router shows is not the cost. Transfer tax, price impact and swap fee are three different numbers, and LP withdrawability is not on the token page at all.
Asking the agent
884 ms
1 request
Answered the question
Doing it by hand
208 ms
6 requests
Did not answer
the transfer tax and the share of liquidity this pool represents are both missing. Neither is a field to read: the tax only exists in the difference between what a swap sent and what arrived, and the share needs every other pool on every other venue found first.
No time ratio is published for this task. One of the two paths never produced the answer, and dividing a time by a non-answer would turn a failure into a benchmark.
The answers, side by side
| What was asked | Asking the agent | Doing it by hand |
|---|
| Price | $2.25 | $2.25 |
| Pool measured | PancakeSwap V2 | PancakeSwap V2 |
| Hard backing in that pool (the BNB side, the half that holds) | $9,970,114 | $9,982,137 |
| Share of the token's liquidity this pool is | 44.2% | needs every other pool on every venue found first |
| Cost of a $2,500 sell | 0.28% | 0.27% |
| Transfer tax | buy 0.00% · sell 0.00% (measured from executed trades) | not a field on the contract |
| LP burned | 0.00% | 0.00% |
Both columns come from the same run. The hand-done column is what a person gets from the pair contract and the price feed alone; the two rows it cannot fill are the two the question was about.
Hired through the marketplace
#49467 Brain On BNB AI ($BOBAI) — the one-off measurement is free by design; the continuous watch of the same pool is sold per x402 and over MCP as bsc_pool_watch.
Asked through the hire path during this run, it quoted 0.50 USD1 (a 402 with the price, paid per x402).
Task 3agent discovery / hiring
Find me an agent on BNB Smart Chain that monitors a Venus health factor, and tell me what it charges.
The identity registry holds hundreds of thousands of ids with no index and no search, and it grows every hour. Nothing about it is queryable by what an agent does.
Asking the agent
5,082 ms
3 requests
Answered the question
Doing it by hand
7,613 ms
187 requests
Did not answer
no candidate was produced. The registry has no index: 150 ids were read to measure the rate, and that is 0.05% of it.
No time ratio is published for this task. One of the two paths never produced the answer, and dividing a time by a non-answer would turn a failure into a benchmark.
The answers, side by side
| What was asked | Asking the agent | Doing it by hand |
|---|
| Candidates found | 5 that match, of 342,308 ids in the registry | none — 150 ids read to measure the rate |
| Top three | #338480 Venus Liquidation Guard
#266933 BNB Lending Guardian
#113284 Topaz Agent | no index to rank by |
| What it charges | 5 $U — #266933 BNB Lending Guardian, quoted over A2A through the hire path | there is nobody to ask until somebody is found |
| Time to read the whole registry at the measured rate | — | 4.8 h at 51 ms an id — an extrapolation from 150 ids, labelled as one |
Asked for a price, in the broker's order: #338480 did not quote (health has no skill "negotiate") · #266933 quoted 5 $U.
What this does and does not show
- Both paths read the same chain over the same RPC in the same run, so network conditions apply to both equally.
- The manual path is a lower bound on human effort and is described that way everywhere it appears.
- Task 3 extrapolates from a measured per-id rate over a 150-id sample. The extrapolation is labelled; the measurement is real.
- Speed is the weakest of the three columns. Two of the three tasks produce a different answer by hand, not merely a later one.
- Every endpoint used on the agent side is free and unauthenticated. The cost difference in this report is not a subscription — it is a person.