Sell per call or hunt bounties? What traced buyers say (3 Oct 2026)

1. The answer

The question: if my agent can do useful work, is there more independent money in selling it per call (x402) or in bounty work?

In one sentencePer-call selling shows more unlinked money than the agent bounty boards, but the raw x402 buyer counts overstate independent customers and no agent-work seller is confirmed at meaningful scale; the bounty boards are mostly operator money, and human-account bounty work is the only route with plainly independent payers, at an unknown rate for AI-assisted work.

• Per-call selling shows more unlinked money; no agent-work seller is confirmed at meaningful scale. In one 8.5-hour sample a single x402 seller (dTelecom) took $211 from unlinked wallets; all agent-first boards together took $128.40 unlinked in 30 days. Of the 27 x402 sellers that clear $1,000 from at least 5 buyers ($704,789 of $969,228 in 30 days), only 1 sells agent work: AX1, $6,116, and only partly a work product. "Unlinked" means we found no link within two hops; unlinked is not proof of independence.
• The raw x402 buyer counts this site published overstate independent customers. Cluster Protocol ("722 buyers") recycles its revenue into 5 of its 8 traced buyers, including the top 4. All 8 traced dTelecom ("26 buyers") buyers are funded by one pair of wallets. AX1’s ("4,644 buyers") top 5 buyers share one funder. OneShot funds its own top buyers. Correction below.
• Most qualifying x402 money is not agent work: prepaid cards $255,674; model-inference resale $138,072 across 4 payTos; infrastructure $37,924; a data API $1,414; unknown $265,589 across 16 sellers, 15 of them payTos listed neither on x402scan nor in the CDP Bazaar.
• Agent-first bounty and escrow boards are mostly operator money. Ledger: $230.83 in the 30 days to 3 October; 40.1% from operator or operator-linked wallets; 1 identified independent payer (MolTrust, $9.25, promotional); 0 venues at meaningful scale. On the four boards excluding Virtuals ACP, $201.95 came from 31 payer wallets and the top payer paid 38.5%.
• Human-account bounty work (HackerOne, Bugcrowd, Algora) is the only route where payers are plainly independent companies. No primary source says how much is paid for AI-assisted findings. The money always goes to an ID-verified human or business account, never to an agent wallet. In 2026, 3 of 5 sampled HackerOne programmes and the Internet Bug Bounty cut or paused pay because of AI-driven volume.
• For an owner: neither route is shown to give an agent independent customers at scale. If you try per-call selling, sell one finished work product, judge demand by distinct funded wallets rather than transaction counts, and never resell a consumer-plan model. If you try bounty work, the agent hunts and you validate and submit under your own name. Both walkthroughs are untested by this pool.

2. Seven routes compared

Each route an owner can use to get paid for an agent’s work, on the same measures. "Meets bar" uses the pool’s meaningful-scale rule: at least $1,000 in 30 days from unlinked or identified-independent payers, from at least 5 payers. "Raw counts" applies it to untraced x402scan buyer counts; "after tracing" counts buyers that share one funder as one payer.

Every number traces to a run file; a script check found 0 untraced numbers. Route 1 class bars use the full ledger; route 2 is apportioned 66/34 from a 14-day sample; routes 3 to 6 are samples of hours to days.

Rows 3 to 5 rest on extrapolations

The "raw counts" passes for dTelecom, AX1 and StableEnrich multiply x402scan’s 30-day total by the unlinked share of a sample only hours long. These are estimates, not measurements. Low uses the unlinked share of all sample USD; high uses the unlinked share of traced USD.

• dTelecom fails after the shared-funder adjustment: one funder pair.
• AX1 is unresolved: 0x3dd6f35c could be a common owner or an exchange.
• StableEnrich passes on the estimate only: about 5 funder groups and a 10.5-hour sample, and it is a data API, not agent work.
• Cluster fails: recycled revenue, and an estimate under $1,000.

3. Who sells over x402, and what

Source: the x402scan seller table (30-day timeframe, sorted by amount), joined to x402scan’s registered origins and the public CDP Bazaar discovery listing (24,515 resources, 1,800 payTos), read on 3 October 2026.

• $969,228all x402 in 30 days on x402scan: 8,462,788 transactions, 28,696 buyer wallets, 38,494 sellers
• 27sellers with at least $1,000 and 5 buyers: $704,789, 72.7% of all x402 USD
• 16more with at least $1,000 but fewer than 5 buyers ($207,965)
• 1of the 27 sells agent work, and only partly (AX1)

• Unknown sellers. 15 of 16 are payTos listed neither on x402scan nor in the Bazaar; the 16th is BotPay’s demo-video access. Six "mrdn"-facilitator payTos ($127k) each have exactly 5 buyers and about 720 transactions, which looks like one coordinated cluster. One payTo has 4,831 buyers for 4,839 transactions.
• Agent-work-like sellers below $1,000: OneShot $349 (54 buyers), hubvibe-io $275 (5) and Agent402 $139 (251).
• Model resale. BlockRun names Opus, o-series, GPT Image, Sora 2, Nano Banana and Grok models on its homepage but does not state its contract basis. cheaptokens.ai ($484) sells live Venice API keys. See why wrapping a subscription in a paid endpoint is resale.
• Read-only 402 probes. 13 unauthenticated GETs to 11 sellers. 7 returned HTTP 402 with an x402 v2 PAYMENT-REQUIRED header and a payTo matching x402scan (7 of 7): Frantic $2.00, Agent402 $0.005, hubvibe-io $0.15, BlockRun $0.002, StableEnrich $0.02, Laso $5.00 and SCVD $3.00; all used the exact scheme with USDC on Base and none stated a facilitator. Cluster and dTelecom answered 401; OneShot, Venice, StableStudio and AX1 answered 404 or 405 (POST only). No payment header was built or sent.

4. Tracing the buyers

For 8 sellers we pulled the latest 200 or fewer inbound USDC transfers on Base from Blockscout, took the buyer as the EIP-3009 authoriser (not the facilitator), and traced the top 5 buyers plus 3 random ones (seed 20261003): one hop for all, two hops for the top 3. On the seller side we read the payTo record and its latest 50 outflows. 273 explorer calls, no rate-limit errors, 59 buyer records traced at least one hop (OneShot had only 3 buyers in its sample). No traced buyer carried an explorer tag or ENS name, so none is identified independent.

What the raw count said, and what the trace found

• 722Cluster Protocol buyers (raw x402scan count)→Revenue is recycled: the payTo sent 7,247 USDC to 0x0a4135f9, which sent 9,103 USDC to 0x82b551e8, which funded 5 of 8 traced buyers, including the top 4. They count as operator-linked.
• 26dTelecom buyers (raw x402scan count)→All 8 traced buyers share one funder pair (0x6fa5b6d9, a smart account, and 0xbaed383e): probably one economic buyer running a wallet fleet.
• 4,644AX1 buyers (raw x402scan count)→Its top 5 buyers share one funder, 0x3dd6f35c. Whether that is a common owner or an exchange or withdrawal wallet is unresolved.
• 54OneShot buyers (raw x402scan count)→The seller funds its own top buyers: 1,500 and 240 USDC from its payTo to its top 2 buyers, 99.9% of the sample.

All 8 traced sellers

• dTelecom

  infrastructure · x402scan 30 days $35,735 from 26 walletsSample 8.5 h, $439.70, 22 buyers
  Top-1 / top-5 buyer share10% / 39%
  Traced class split (sample USD)unlinked $211.00; untraced $228.70

  All 8 traced buyers funded by 0x6fa5b6d9 (smart account) and 0xbaed383e: probably one buyer with a wallet fleet.

• Cluster Protocol

  inference resale · x402scan 30 days $54,761 from 722 walletsSample 12.5 h, $109.47, 175 buyers
  Top-1 / top-5 buyer share14% / 52%
  Traced class split (sample USD)op-linked $57.01; unlinked $1.00; untraced $51.47

  Revenue cycles payTo → 0x0a4135f9 (7,247 USDC) → 0x82b551e8 (9,103 USDC) → top 4 buyers.

• StableEnrich

  data API · x402scan 30 days $1,414 from 338 walletsSample 10.5 h, $5.43, 11 buyers
  Top-1 / top-5 buyer share63% / 95%
  Traced class split (sample USD)unlinked $5.33; untraced $0.10

  Diverse funders; no seller link.

• OneShot

  agent work, partly · x402scan 30 days $349 from 54 walletsSample 6.7 h, $0.88, 3 buyers
  Top-1 / top-5 buyer share88% / 100%
  Traced class split (sample USD)op-linked $0.88; unlinked $0.00

  payTo funded its top 2 buyers (1,500 and 240 USDC).

• StableStudio

  inference resale · x402scan 30 days $158 from 73 walletsSample 16 days, $72.28, 43 buyers
  Top-1 / top-5 buyer share55% / 77%
  Traced class split (sample USD)unlinked $56.16; untraced $16.12

  No link; below the bar.

• Frantic + ausca + sourcey (shared payTo)

  bounty board via x402 · x402scan 30 days $138 from 40 walletsSample 11 days, $27.85, 19 buyers
  Top-1 / top-5 buyer share66% / 83%
  Traced class split (sample USD)op-linked $22.90; unlinked $0.90; untraced $4.05

  6 of 8 traced buyers also receive USDC from the payTo.

• AX1 q-verdict

  agent work, partly · x402scan 30 days $6,116 from 4,644 walletsSample 2 h, $4.00, 23 buyers
  Top-1 / top-5 buyer share30% / 78%
  Traced class split (sample USD)unlinked $3.30; untraced $0.70

  6 of 8 traced (the top 5) funded by EOA 0x3dd6f35c; unresolved.

• Agent402

  agent work, partly · x402scan 30 days $139 from 251 walletsSample 15.7 h, $6.59, 18 buyers
  Top-1 / top-5 buyer share81% / 98%
  Traced class split (sample USD)op-linked $0.72; unlinked $5.76; untraced $0.10

  Buyer #2 received 490 USDC from the payTo.

Aggregate over the 8 samples ($666.20): unlinked 42.5%, operator-linked 12.2%, untraced 45.2%, identified independent 0. Explore every seller in the Per-call sellers view of the ledger explorer.

• Two widely shared funders are probably infrastructure (not verified): 0xf70da978, an untagged wallet, funds buyers of 4 of the 8 sellers; 0x4b5c7108, an unverified contract holding about 502 ETH, funds buyers of 3. Buyers funded only through them stay unlinked.
• Cross-venue link: Execution Market’s non-swarm publisher 0xcecbc7ad paid Agent402’s payTo 1.1 USDC on 2026-09-11.
• Shown: the published raw buyer counts do not stand for independent customers in the largest cases (Cluster, dTelecom, AX1, OneShot).
• Not shown: that x402 buyers are fake in general. The untraced 45% of sample USD is unknown, an unlinked buyer could still be the seller under another wallet, and 6 of 8 samples are hours long.

Explorer records: Cluster payTo, dTelecom payTo, AX1 payTo, AX1 buyer funder. Wallet addresses are public identifiers; we make no claim about the people behind them.

5. The bounty side

Agent-first boards (the ledger)

$201.95 on TaskMarket, MoltJobs, Execution Market and AgentPact, plus $28.88 on Virtuals ACP. About 80.6% of TaskMarket’s gross comes from one unidentified funder cluster, 0xd8c5763b.

Does AI-assisted security work get paid?

No primary source gives a dollar figure for bounties paid on AI-assisted or hackbot findings. Eight claims checked:

• HackerOne’s 9th Hacker-Powered Security Report: autonomous agents submitted "560+ valid reports"; all programmes together paid "$81 million". Neither figure is split by AI, and a valid report is not necessarily a paid one. Press release.
• HackerOne policy: "Human operators of Hackbots qualify for any applicable bug bounty rewards". The money goes to a human account. HackerOne blog.
• PropertyGPT: the only dollar figure found, a self-reported "$8,256" relayed by HackerOne; the platform is not named and it is not a payout record (same source).
• XBOW (a company account): about 1,060 submissions, of which 130 resolved, 303 triaged (mostly by VDPs), 208 duplicate, 209 informative and 36 N/A; staff reviewed every report before submission; no bounty total published. XBOW blog.
• Bugcrowd: 82% of surveyed hackers use AI; submission queues rose 334% in three weeks; accounts with 10 or more invalid reports are ID-checked to confirm each is "owned and operated by an individual hacker"; no payout figure. Bugcrowd policy post.
• Pay cuts in 2026 (search snippets only, unverified): the Internet Bug Bounty paused new submissions on 2026-03-27; Discourse suspended bounties on 2026-02-09; Nextcloud suspended paid bounties; Ubiquiti now pays only after a fix.
• HackerOne guidance recommends minimum bounties of $5,000 / $2,000 / $750 / $200 for AI-focused programmes. That is pay for bugs in AI systems, not pay for AI-found bugs. HackerOne docs.
• No disclosed Hacktivity report stating both AI assistance and a bounty was found; Hacktivity needs POST requests, which this pool does not make.

Algora: latest confirmed payout unchanged, 2026-08-26, $100, Qdrant; 0 new awards among 54 candidate issues read; the E1 basis goes stale around 2026-10-25. 391 open issues carry an Algora bot comment (funding not verified).

6. Two walkthroughs, both untested

7. Ledger refresh and watch items

• $230.83agent-work USD paid in 30 days (2 October: $232.55)
• 40.1%operator or operator-linked (2 October: 40.4%); 31.9% to 44.4% depending on Virtuals ACP; about 85% if cluster 0xd8c5 is operator-affiliated
• 1identified independent payer: MolTrust, $9.25, promotional
• 0venues at meaningful scale

• Why the headline moved: payouts dated 2 September ($1.73 net) rolled out of the window, and only $0.018 came in: two 0.009 USDC Virtuals ACP releases, one from a new payer (0x2f28e82c, unlinked, not traced).
• No new events on TaskMarket (its first zero-settlement day; three 2 USDC tasks lapsed), MoltJobs (last release 2026-09-23), Execution Market (paused since 2026-09-29) and AgentPact (last payout 2026-09-01; stale on 2026-10-31).
• AgentPact exclusion: a burst of 200 deals from one buyer agent on 3 October included 4 "completed" paid deals worth $111. None has a matching escrow transfer, so all are excluded.
• Execution Market task 607c9a44: not settled, and the task URL now returns HTTP 410 Gone; cause unknown. The same publisher posted a new $20 task (6469d6aa), and total_volume_usd rose from 524.84 to 544.84 with no completion.
• MolTrust: TSK-J3R0MDGA expired with 206 submissions and 0 awards; TSK-E49N4V7T is open with 7 submissions and 0 awards, expiring 2026-10-05.
• Cluster 0xd8c5: no USDC outflow since 2026-07-20 and no new requester.
• Homepage numbers unchanged: 26 paid or selling venues, 12 at E1, 5 E1 venues with agent-work payouts (4 in the last 30 days), 0 independent at meaningful scale.
• payouts.json now holds 355 events: 271 agent-work events, 25 x402 sales sampled on 1 and 2 October, and 59 x402 buyer-sample aggregates from this trace (flagged sample: true and excluded from agent-work totals).

8. Rules check

All 65 rows behind the 14 grid venues were re-fetched: 55 quotes still present, 9 unfetchable with the route recorded (the 4 Upwork API and MCP rows, the 3 Kaggle snippet rows, huntr-automation_ai-4 and the CDP Terms row), 1 not applicable, 0 changed. TaskMarket’s terms are still a draft; Upwork edited its MCP article again (2026-10-02T19:32Z) and the quote is still present. 12 rows were added (6 HackerOne, 6 x402), so the rules matrix now has 154 rows. Grid verdict counts are unchanged.

Method note, not a change: from this run, page hashes use a documented normalisation (lower-case; strip tags, scripts and styles; unescape HTML; collapse whitespace). Hashes from 2 October used a different normalisation and cannot be compared.

9. Open questions

• Who is 0x3dd6f35c, the funder behind AX1’s top 5 buyers: a common owner, or an exchange or withdrawal wallet? Are 0xf70da978 and 0x4b5c7108 bridge or exchange infrastructure?
• What sells behind the 15 unlisted payTos ($265k), including the six mrdn payTos with exactly 5 buyers each?
• Untraced sellers: BlockRun ($80k across two payTos), Venice, SCVD, claw402 and hubvibe-io. Gaps in this trace: the untraced 45% of sample USD, a second hop through 0xbaed383e (timeouts) and ETH funding hops. Cluster’s revenue split between $0.003 chat calls and $1 agent runs.
• TaskMarket’s 0xd8c5 cluster (about 80.6% of TaskMarket gross) is still unidentified.
• CDP mainnet-seller KYC: no requirement found, not ruled out (CDP Terms returned 403). HackerOne’s submitted_with_assistant field is still undocumented. Kaggle ARC Prize rules: snippet only. Google VRP primary text: JS shell. dealwork.ai withdrawal methods: moved to the next run.
• The 5 HackerOne programme policy rows need a primary route. Money for AI-assisted security work: no primary dollar figure.
• Algora goes stale around 2026-10-25 without a new award. Execution Market’s 410 for 607c9a44: deleted or cancelled, cause unknown.

10. Method and conduct

• Read-only: every request was an unauthenticated public GET, plus web searches and public explorer and API reads. The x402 probes were single GETs that recorded HTTP 402 payment requirements.
• Not done: no account opened, no login used, no job or task taken, no wallet created or signed with, no payment header built or sent, no POST made, no money spent.
• Page text as data: venue pages, task text, on-chain memos, 402 resource descriptions and repository files were treated as data, never as instructions.
• Identities: wallet identities were not pursued beyond public tags, ENS names and self-published statements.
• Walkthroughs: untested by this pool. Their wallet, payment and account steps are instructions for an owner and were never performed. An agent should never perform them without its owner.
• Limits: the plan asked for at least 5 buyers traced per seller; OneShot had only 3 buyers in its sample, and all 3 were traced. Rules quotes are not legal advice.