Does an agent beat doing it yourself?
Three questions a DeFi user actually asks, answered once by an agent and once by reading the chain directly, then timed and checked against each other. Measured 1 September 2026.
Setup
- Agent
- HeyAnon ERC-8004 Venus agent
- Identities
- #43129, #45381
- Endpoint
- https://erc8004.heyanon.ai/mcp/venus
- Availability
- 14/14 probes answered, measured by Pokter
- By hand
- Direct contract reads over a public BNB Chain RPC, written for each question
Timing is wall-clock for one run each. A single run says something about shape, not a distribution; nothing here is a benchmark.
Venus supply APR
What is the current Venus supply APR for USDT, USDC and BTCB on BNB Chain?
USDT: 2.83%, USDC: 2.15%, BTCB: 0.19%
One tool call.
vUSDT 2.84%, vUSDC 2.15%, vBTC 0.19%
Comptroller market scan, 55 markets, symbol resolution, annualisation.
Quality. Equivalent. Two independent paths produced the same three figures to within rounding, which is what verifies the agent — not anything it claims about itself.
Winner: AgentSame answer, far less setup, and it avoids the annualisation trap.
Liquidation risk on a live position
Is 0x96145D06…8Bf1, a real Venus borrower carrying ~1.96M USDT of debt, close to liquidation?
borrowLimit $1,738,691.15 · shortfall $0.00
One tool call.
liquidity $1,736,364.45 · shortfall $0
A single Comptroller.getAccountLiquidity read.
Quality. Same decision — no shortfall, not liquidatable — but the headroom differs by about $2,327 (0.13%). Likely different blocks with oracle prices moving between them; we did not confirm that and do not report it as confirmed.
Winner: By handWhere the primitive already exists and you know its name, wrapping it in an agent adds latency and nothing else. Manual won by 3.3×.
Borrow cost across chains
What does it cost to borrow USDT and USDC on Venus, on BNB Chain versus Ethereum, to price the borrow leg of a carry trade?
BSC — USDT 4.54%, USDC 3.95% · Ethereum — USDT 3.92%, USDC 7.65%
Two chains, four markets, one call.
BSC — USDT 4.54%, USDC 3.95%
One chain only. Ethereum needs a second RPC and a second market scan.
Quality. Identical where they overlap; the agent covers strictly more. The answer is only interesting because it spans chains — USDT is cheaper to borrow on Ethereum, while USDC is nearly twice as expensive.
Winner: AgentGreater coverage for the same wall-clock budget.
The annualisation trap
Venus reports rates per block, and the conversion to a yearly figure depends on how many blocks BNB Chain produces in a year. That number has changed. At the measured 0.45 s block time there are about 70,080,000 blocks a year. Code still using the older 10,512,000 reports 0.43% where the true figure is 2.84%: off by 6.67×.
The agent answered 2.83%. It did not fall in. A hand-written script easily does, which is a real point in the agent’s favour that a stopwatch would not show.
How to read this
An agent earns its fee where it saves setup, covers more ground, or avoids a mistake you would make. It does not, where the primitive exists and you know its name: there it adds latency and a party to trust. That is why one of the three goes to doing it by hand.
This is one agent, one run per task, on one date. Pokter’s ongoing evidence about any agent is on its page and in the methodology. The full write-up is in the repository.