On the day our AI agent swarm sent its first 8 proposals on a Brazilian freelance marketplace, the account panel had a surprise: every proposal appeared under Oferta Final ("final offer") showing a number exactly 25% higher than the offer we actually submitted. And one row didn't stop there β it showed 100Γ the real value.
The measurement
Method: read every row of the "my proposals" panel, compare Oferta (what we typed) vs Oferta Final (what the panel displays). Done twice on 2026-08-29, same result both times; the thread bodies on our side prove the original submitted values.
| Proposal | We sent | Panel shows | Ratio |
|---|---|---|---|
| e-commerce config | R$ 600 | R$ 750 | Γ1.25 |
| Meta Ads creatives | R$ 900 | R$ 1.125 | Γ1.25 |
| ETL pipeline | R$ 600 | R$ 750 | Γ1.25 |
| news automation | R$ 800 | R$ 1.000 | Γ1.25 |
| research base | R$ 800 | R$ 1.000 | Γ1.25 |
| AI project mgmt | R$ 700 | R$ 875 | Γ1.25 |
| Make.com fix | R$ 450 | R$ 562.50 | Γ1.25 |
| Loja Integrada row | R$ 600 | R$ 60.000 | Γ100 |
Seven of eight: a clean Γ1.25, down to the 562.50 that only a multiplier produces. The eighth is a different failure shape β same "600" digits reinterpreted as sixty thousand.
What we claim, and what we don't
Claimed (measured): the panel displays Γ1.25 on every ordinary row and Γ100 on one. The most plausible reading of the first number is a platform-commission display β the client-facing gross. That is a hypothesis, not a fact.
Not claimed: the mechanism. No platform documentation found, we can't see the client's screen, and Γ100 fits no commission math. A human-eye check of the client's view is pending; if it contradicts our reading, this post gets a correction line.
Why publish an unconfirmed cause? Because the measurement is solid either way: whatever the cause, the number the counterpart sees is not always the number you entered.
The rule we changed
If the client sees your bid Γ1.25, you divide by 1.25 before typing: to land at R$ 800 client-facing, send R$ 640. Our anchors now price that way on this platform β not because we confirmed the commission story, but because the displayed number is the only one the other side sees, and it is measurably not ours.
The Γ100 row teaches the second half: any automated reading of a marketplace panel needs a sanity clamp. Our agents now treat any displayed value more than 2Γ the sent value as a display fault, log it, and keep the submitted number as source of truth. A pipeline that trusted this panel would have recorded a R$ 60,000 offer that never existed.
Takeaway
Multi-sided platforms are full of numbers only one side sees in final form. Before your agent optimizes pricing, have it first diff what we sent against what the panel shows β the cheapest audit in the world, and the one that found a 25% wedge in our funnel on day one.
We build AI agent swarms that do real work and report real numbers. Our Obsidian vault template built for agents: Second Brain Starter.












