Does an agent actually beat doing it yourself?

Three tasks, each run twice against live BNB Chain state on 2026-09-01: once through a live third-party agent, once from primary sources with no agent involved. The agent won 2 of 3 — and the one it lost is the more useful result.

Agent under test: HeyAnon ERC-8004 Venus agent
ERC-8004 ids: #43129, #45381
Availability: 14/14 probes answered, measured by Pokter

The finding: it wins on correctness, not speed

Venus publishes a per-block rate. Annualising it needs a blocks-per-year constant, and the figure in most documentation assumes BSC's original three-second blocks. It no longer produces them.

Measured blocks/year: 70,080,000; 0.45s blocks, read from chain
Published constant: 10,512,000; assumes 3s blocks
Resulting error: 6.67×; 0.43% instead of 2.84%

An analyst copying the published constant computes 0.43% where the true figure is 2.84% — no error, no warning, just a number wrong enough to make a profitable position look not worth entering. The agent answered 2.83%.

The experiments

Yield optimisation

Venus supply APR

What is the current Venus supply APR for USDT, USDC and BTCB on BNB Chain?

Agent wins

Agent2.28s

USDT: 2.83%, USDC: 2.15%, BTCB: 0.19%

One tool call.

Manual3.23s

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.

Verdict: Same answer, far less setup, and it avoids the annualisation trap.

Health-factor monitoring

Liquidation risk on a live position

Is 0x96145D06…8Bf1, a real Venus borrower carrying ~1.96M USDT of debt, close to liquidation?

Manual wins

Agent2.93s

borrowLimit $1,738,691.15 · shortfall $0.00

One tool call.

Manual0.90s

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.

Verdict: Where the primitive already exists and you know its name, wrapping it in an agent adds latency and nothing else. Manual won by 3.3×.

Grid / trading · satisfies the Trading requirement

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?

Agent wins

Agent2.46s

BSC — USDT 4.54%, USDC 3.95% · Ethereum — USDT 3.92%, USDC 7.65%

Two chains, four markets, one call.

Manual2.45s

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.

Verdict: Greater coverage for the same wall-clock budget.

How this was measured

Agent calls are plain JSON-RPC against a read-only MCP server; no funds moved. The manual baseline reads the Venus Comptroller and per-market rate accessors directly.

On “manual time”: the baseline was performed by an AI assistant working from primary sources, not by a human analyst. The timings are real wall-clock measurements, but they are machine-to-machine — a person would be slower on the manual side, so where manual loses it loses by more than shown. We would rather name that limit than invent a human figure to fill the column.