The Agent Labour Market, Probed: A Complete Bounty Board With Four Customers
Can an agent take a job, do it, and get paid, with no human anywhere? The mechanism exists, it is complete, and it speaks x402. Salesmart S.r.l., which publishes Sinapsi, called it directly and enumerated the discovery index around it. The economy on top of it does not exist yet: the board took 43 requests in thirty days from four distinct wallets.
There is no registration. The wallet is the account.
The first surprise is what is missing. No signup, no email, no profile, no identity document. The credit service states the rule in one line: "Balance is keyed to your paying wallet." Pay once and the platform knows you.
An agent can therefore be a customer and a worker here without a human ever standing behind it: the only credential is a wallet that has paid. We had written a comparison with the classic freelance marketplaces at this point — that their terms require a human identity and a bank account and forbid automation. We could not source it: upwork.com/legal and fiverr.com/terms_of_service both answer HTTP 403 to WebFetch *and* to curl with a browser user-agent (tried 2026-08-05), and the claim rested only on a web-search summary. It is cut, not softened. Nothing in this page depends on it.
The full labour cycle, in the operator's own words
x402.shizu.me/bounty calls itself "Agent Bounty Board" and tags itself
bounty, marketplace, tasks, labor, agent-commerce. Its declared mechanism, quoted verbatim
from the payload it serves:
> "POST {spec, reward_credits} escrows the reward and lists the task; GET lists open > bounties, GET ?id=X reads one task (any state) incl. its full spec. Workers claim > (/bounty/claim), submit results (/bounty/submit); the poster approves (/bounty/approve) > and the worker is credited minus a 20% house fee. TO POST: hold reward_credits credits > first (POST /credits/buy, $0.010 each); short balance answers 409."
Every stage of an employment relationship is there — posting, escrow, claim, delivery,
approval, payment, and a house cut. Reading the board costs $0.005 in USDC on Base
(eip155:8453), paid to 0x217e5Fe265EB78b29067bF8324ef03a7D8e167C4. The internal unit is a
credit worth $0.001.
Note what the escrow implies: someone must approve the delivered work before the worker is paid. That is the missing conduct layer of x402 reappearing as a product feature — the protocol authorises a transaction and is blind to whether the work was any good (A Spending Limit Is Not a Conduct Limit).
The numbers: a complete machine, almost unused
| | | |---|---| | Bounty board requests, 30 days | 43 | | Distinct paying wallets | 4 | | Credit-balance checks / top-ups | 33 / 41 | | Whole labour-and-trust slice (495 listings) | 2,680 calls — 0.70% of the index |
The 495-listing slice is a crude keyword cut (bounty, escrow, task market, hire,
gig, freelance, reputation, trust) and over-collects: "most trusted marketplace" in a
scraping vendor's blurb matches it. Declared as a rough upper bound, not a clean category.
The three figures are produced by agent-economy/dati/slice_labour.py, which is committed
next to the snapshot and states in code which fields it matches on, so anyone can rerun the
cut and get the same numbers. On the 5 August snapshot the same script returns 496 listings
and 2,783 calls, 0.74%.
Against the index total we measured the day before — 382,225 calls in 30 days across 14,766 listings (The x402 Market, Counted: 14,766 Listings, 12,741 Calls a Day, One Winner) — agent labour is a rounding error.
One listed platform is already dead
x402.devvizion.com/v0/bounty-radar, described in the index as a "multi-source open bounty
radar", answers 404. It is listed, priced, and gone. This is the churn that the trust
services below exist to report, observed directly rather than inferred.
The real business in this slice is doubt
Sorting the slice by usage, most of it is not labour. It is services selling an answer to
*"can I trust that other service?"* — trust scores for x402 endpoints, wallet-reputation
oracles, firewalls that scan an incoming 402 offer before you pay it, and a "service
graveyard" listing endpoints that vanished from the census. AgentPay
(agentpay.tools/tools/verified_route/call) describes itself as a "Buyer-side trust oracle
for the x402 marketplace: 'I need X, budget $Y — which tool is real?'" — *real*, not
*suitable*: the question the buyer needs answering is whether the counterparty exists at all
and is not one wallet stamping many fake-distinct listings. Truth Bear maintains a trust
roster of 1,601 services.
A whole cottage industry has grown to answer the question the discovery index does not, which is independent confirmation that the gap is real and felt.
But look at what they can measure: liveness, uptime, whether the wallet looks organic,
whether the service still exists. Reading the descriptions of all 495 listings in the
3 August 2026 slice, we found exactly one that claims to judge the delivered result
rather than the seller: verifier.vevang.com/verify, "AI Output Verifier — independently
verify another agent's deliverable meets its spec BEFORE you pay for it … judges
spec-satisfaction". It priced that at $1.00 and took one call from one wallet in thirty
days. (A second, api.relaystation.ai/v1/contracts, sells "paid arbitration when disputes
arise": 23 calls, 2 payers.) So the correction is not that outcome-verification is
impossible — someone is selling it — but that nobody is buying it, while seller-reputation
services in the same slice are the busiest thing in it. That is the
difference between reputation-of-the-seller and verified-usefulness-of-the-content
(Four Ways to Sell Content to an AI, and Who Keeps What, The Earn-Spend Loop: Why Machine Payment Is Half-Built).
The largest earner in the slice is none of these: it is Apify selling prepaid tokens for its scraping marketplace at $1.00, with 149 calls from 30 payers — an ordinary tools business, not an agent-labour one.
A protocol detail worth having seen
The board answers in both protocol versions at once: the response body declares
x402Version: 1 with an X-PAYMENT header and the legacy network string "base", while the
payment-required response header carries the v2 payload with the CAIP-2 identifier
eip155:8453. Anyone implementing a seller must accept both, and here that requirement is
visible in the field rather than inferred from the specification
(x402: HTTP 402 Finally Gets a Job, at 32 Cents a Transaction).
What this changes, and what it does not
It does not show an agent labour market. Forty-three reads a month by four wallets is a demonstration, not an economy.
It does settle the question of whether the machinery exists: it does, end to end, including escrow and an approval step, with the wallet as the only identity. Nothing needs to be invented for an agent to earn — only customers need to appear.
Falsifier. We are wrong to call this negligible if, by 30 June 2027, any agent-labour
board reports four figures of monthly settled task value, or the distinct-payer count on one
of them exceeds 100. Both are checkable with the same probe that produced this page; the
snapshot script is in agent-economy/dati/.
Verified against
16 claims checked against these sources · 3 refuted and removed
What links here
Source: Sinapsi — verified compositional memory, queryable by LLMs. Query this wiki live from your assistant over MCP, or build your own verified wiki (public, or private for your team). CC BY 4.0 — reuse with attribution to Sinapsi.