We funded the wrong wallet first. Then an AI agent paid a developer 0.1 SOL.
We sent about 0.84 SOL to the wrong wallet before the agent had anything to spend. I had reported the wallet I sign in with as the agent's wallet. Steve Capital had created a different one.
That was the first of several places where Steve Capital, OOBE Protocol's Solana agent product, did something other than what we assumed. We entered its Agent Challenge on Superteam Earn with an agent called Milestone Pay. It pays one real developer a retroactive 0.1 SOL for a merged pull request in a public Solana repository, and it should refuse a payment that breaks its risk limits. Everything below ran on Solana mainnet on 18 November 2026.
The X thread has the short version and the YouTube walkthrough has the screens. This post has the failures. An AI assistant drove my signed-in browser for most steps, and I approved each spend.
The wallet we signed in with wasn't the agent's wallet
Steve creates a separate agent wallet at sign-in. Our Solflare wallet signs us in, and the agent wallet holds the funds. I funded Solflare with 0.843331095 SOL. Steve's public profile shows the agent wallet as 9R9hkm…Q2D5jB, and mainnet RPC showed no account at that address.
The bounty needs the agent wallet to hold at least 20 USDC or equivalent SOL. We sent it 0.36 SOL from Solflare, about $27 at $75.08 per SOL. Before the payout I added 0.06 SOL to keep it above $20.
A 0.05 SOL tool cap would have blocked our own payout
We tightened Steve's wallet risk limits from the defaults.
| Setting | Steve default | Ours |
|---|---|---|
| Per transaction | 0.5 SOL | 0.1 SOL |
| Max position | $250 | $25 |
| Daily loss | $50 | $10 |
| Max leverage | 3x | 2x |
| Per tool | 0.05 SOL, 25 USDC | 0.05 SOL, 15 USDC |
The per-tool cap then got in our way. A MultiHopper route like ours needs 0.079457414 SOL up front, which is above 0.05, and the cap would also have stopped a 0.1 SOL transfer. For the payout we raised the per-transaction and per-tool SOL caps to 0.15, then put them back to 0.1 and 0.05.
The chat model said it had registered us. No tool had run.
Steve's Arena has a mission to register the agent in the SAP Agent Registry, an on-chain registration. We asked Steve chat to do it. The first call to sap_build_agent_register_transaction failed with "Non-base58 character", because the free model passed a bad wallet argument. The next two replies claimed a ready transaction with an agent address. No tool ran and no signing prompt appeared. We checked mainnet: that address doesn't exist.
With the wallet passed explicitly, Steve's get bridge status tool reported onchainAgentRegistrationEnabled: false, sapPaymentsBridgeEnabled: false and localSigner: blocked. Registration is switched off. The default sponsored model is cohere/north-mini-code:free.
At payout time nemotron-3-ultra reported a "quote-only swap preview" blocker that no tool call showed. We now verify on-chain before trusting a chat claim. We asked OOBE on Discord whether the registry mission is disabled; no reply is recorded.
MultiHopper quoted to the lamport, and quoting was all Steve could do
MultiHopper routes a payment through a series of hops and returns unsigned transactions. We created an API key and asked it to estimate a 10 USDC, 3-hop route. The fee was 0.049751 USDC (0.5%), the recipient would get 9.950248 USDC, and 0.079457414 SOL was needed up front.
Inside Steve, External MCP · estimate transfer returned the same figures: fee 49,751, delivered 9,950,248, requiredSolUpFrontLamports 79,457,414. But Steve's built-in MultiHopper provider has two tools, list the hop library and estimate a transfer. It uses a Steve-managed key and has no place to add ours, so Steve cannot create, prepare or confirm a route.
The MultiHopper track requires transactions through the Steve agent. A route made with our own key outside Steve would count as our integration's, not Steve's. So the payout did not go through MultiHopper, and we aren't claiming that track. We asked OOBE whether create and prepare tools are coming before the deadline. No reply is recorded.
A dedicated Ace app was the only way to show the credits were bought
The Ace Data Cloud track needs at least five successful paid requests on purchased credits. Free credits don't count. Our console had 2.0 trial credits, which would cover five searches at about 0.05. Credits bought into that general balance would mix with them, and the usage rows couldn't show which paid.
So we created a dedicated Search Engine application. It came with 0.2 free credits of its own. We switched off "Allow Use General Balance" and bought the smallest package, 40 credits for $7.00 as of 18 November 2026. The balance went from 0.2 to 40.2. Each search cost 0.01 credits, and the balance after deduction fell from 40.19 to 40.15, with a trace ID on every call. The order page still said "Continue Payment" after we paid in crypto, so the balance is the evidence, with the order receipt.
The five searches fed a review brief. No result named the repository. Search informed the decision. It didn't authorise the payment.
We paid a colleague, and his PR has weaknesses
MP-001 pays one real, independent person. Synthetic personas appear only as labelled refusal test cases, never as payees. We posted a call in a Telegram community group, but the contributor is a colleague I asked directly. That departs from the "first eligible reply" rule we posted.
His PR is #1 in cirko33/master-multisig, a public Anchor multisig: one file, +376 lines, opened on 18 October 2025, so it predates the challenge. He merged it himself, in his own repository, 10 seconds after opening it. The repository has no licence.
We paid because the PR is real, public, Solana-specific, written by the wallet's owner and not made for this challenge. We sized the reward to match. He posted his wallet from his GitHub account in a PR comment and consented in writing there to being shown publicly. No wallet match, no payment. That is his acceptance of his own work, not an independent maintainer's.
The refusal appeared in chat, not in the form
To get a refusal on camera we asked for 0.2 SOL, double the 0.1 SOL cap. In the Portfolio transfer form, three tries went the same way. "Build transfer preview" closed the panel with no message and created nothing. The cap held, but a viewer couldn't tell why.
In Steve chat the same request ran steve_get_wallet_risk_policy (5 ms), which returned the 0.1 SOL cap. The agent refused, quoting "maximum of 0.1 SOL per transaction" against the 0.2 SOL requested, and offered a 0.1 SOL transfer. No transaction was built. The silent form is on our friction list for OOBE.
The transfer form said "OOBE fee 0 bps" and still took a fee
Chat couldn't build the payout. The default model hit its step budget twice, and the stronger one claimed the blocker above. The Portfolio transfer form built it: 0.1 SOL to the contributor, "Preview passed", and a Steve protocol fee of 0.007190754 SOL ($0.50 fixed plus 0.5%), despite the form's label. The preview included the fee before we approved.
The transaction landed at 08:47:46 UTC: 3WcDaFQr…zaXQi. It has two System-program transfers, 0.1 SOL to the contributor and the fee to Steve, plus a 5,000-lamport network fee. The fee was 7.19% of the payout.
The video showed its captions twice
The first renders burned captions into the picture and shipped a same-named .srt file beside it. Desktop players load that file automatically, so viewers saw both. The 28 cues themselves had no overlaps. Version 3 has no burned-in captions, at most two lines of 42 characters per cue, and a separate caption track on YouTube.
The narration is a synthetic voice (Kokoro, preset af_heart, generated locally). The screens are real captures.
How much to trust this
This is one payout to one recipient on one day, with five paid searches behind it. The funding, payout and fee are on-chain, and you can check them. What Steve did inside its own tools comes from our screens and log.
The recipient is a colleague, his PR was self-merged in an unlicensed repository, and the project is small. SAP registration and any MultiHopper transfer did not happen. OOBE's own availability table lists Steve Capital OS as "product development", so behaviour and prices here are as of 18 November 2026.
Steve's settings showed the agent wallet's key available for local signing on this device, with the recovery vault not set up.
What we'd keep
The pattern: research, policy checks, human-approved payment, receipt. Check the agent wallet's address on its public profile before funding it. Set limits to what the task needs, then ask the agent to break them. Check the chain for any claim that should be on it.
Receipts: the payout, the video and the thread. If you run an agent with its own wallet, what would you want it to check before it pays someone?
The SOL, fees and Ace credits in this experiment came out of our own pocket. If you'd like to help fund the next one, you can buy us a coffee.