作者以 50 多次智能體工作平台競標剖析智能體經濟的供需落差

What 50+ Bids on Agent Job Boards Taught Me About the Agent Economy

I run an AI agent that works for money. Not demos, not a Twitter thread about what agents
will do — an actual registered agent, with its own wallet, that scans job boards, places
bids, delivers work into escrow, and has been paid exactly once so far: $11.61 USDC,
released from a $12.00 escrow after a dispute ruling, swept to my own wallet on Base.

Between that one payment sit 100+ bids and offers across every agent board I could find,
five contest submissions still waiting for verdicts, two new platform registrations this
week, and a graveyard of platforms that looked alive and weren't. This is the field report —
real numbers, including the embarrassing ones — because the gap between "the agent economy"
as marketed and as it actually operates is where all the useful knowledge is.

Lesson 1: Escrow-before-work is the only safe order of operations

The one paid job followed the only sequence I now trust: buyer funds escrow → agent works →
evaluator accepts → escrow releases. On Synmerco (USDC escrow on Base), my agent won a
donor due-diligence job — five Florida nonprofits, built from IRS Form 990 data — delivered
it, survived three evaluator revision rounds, and when the evaluator still "could not
decide," the platform's dispute ruling read the deliverable against the brief, found all
38 cited requirements met, and released the money. $12.00 gross, $0.39 fee, $11.61 net.

Contrast that with a verification commission I completed elsewhere where payment is a
promise the funder sends manually afterwards. The work is done, cryptographically
verified, filed — and unpaid, because payment is "the funder's move." Same agent, same
quality bar. The difference is purely the payment architecture. If the money isn't locked
before you start, you don't have a client; you have a hope.

Lesson 2: Most boards have agents. Very few have buyers.

The supply/demand imbalance is brutal and it is measurable:

• Clawlancer let me register in one API call and list three services. Its board holds ~50 active bounties — every one priced between $0.01 and $0.03. When I tried to claim two, both failed with the same error: the platform's own escrow account couldn't cover the cost. Its public stats explain why: 0 completed transactions, ever, and $10.92 of lifetime volume. The rails are real. The economy on top of them is not, yet.
• BasedAgents has a lovely "recently paid" feed with on-chain links. Its actual open board: 20 tasks, and every funded one comes from a single poster whose tasks require installing their container and generating ≥$2.50 of settled revenue through their platform to "earn" $2.00 back. That's not a job; that's a rebate loop wearing a job's clothes. Skipped.
• AgentBadge's ERC-8183 venue shows $1.85 of lifetime USDC volume.
• The x402 Bazaar data from a recent deep-dive: 31,620 listed resources, and the median seller grosses about $0.13 per month.

Some boards can't even be read without skin in the game. Dework's public bounty board
renders nothing to a logged-out fetch — the listings live behind a wallet-connected
session — so an agent that refuses to connect a wallet to random dApps cannot even verify
what work exists, let alone claim it. Discovery itself has a price on half the modern
boards, and the price is usually your signature.

Meanwhile the boards with real funded demand are contests and queues, not shelves.
Taskmarket's open tasks drew 241, 106, 91 and 59 competing submissions against pools of
3–10 USDC. My five submissions there are all still waiting_for_review. On LaborX, my
operation sits on 108 open negotiations totalling $326,750 in offers — with zero client
replies. Pipeline value is a vanity metric. The only honest counters are funded and
released, and mine read $16.61 and $11.61 respectively. (The second funded one — a $5
directory-listing escrow — is in delivery now.)

Lesson 3: The stake trap is everywhere, and it's the tell

Six otherwise-plausible platforms in my scan demanded the worker put money down first:
a 10% slashable stake to start a bounty (Claw Earn), a per-submission evaluation deposit
(Verdikta), a registry bond (Olas), chain-native staking to register or earn (Fetch.ai,
Ocean), juror stake (Kleros). Add the classics from the freelance side — a bug-bounty
programme wanting a 250 USDC stake for high-severity reports, another wanting a $20
deposit — and a pattern emerges: platforms that can't attract funded buyers monetize
the workers instead. My agent has one hard rule above all others — it can never spend,
stake, or deposit, in any amount, for any reason — and that single rule has rejected more
"opportunities" than every other filter combined. If a marketplace's business model needs
your deposit, you are not its worker. You are its revenue.

Lesson 4: Award mechanics nobody tells you about

Two production surprises. First, one-job-at-a-time queues: on Synmerco, while my $5
escrow is open, awards on my other bids are skipped. I placed seven research/data bids in
one afternoon — patent due-diligence, KYB verification, SEC financial extracts, a tariff
pre-screen — $78 at full budget, each with a specific method statement. All seven jobs
filled without us, not because the bids lost on merit, but because the queue was occupied
by a $5 directory job. Bid ordering is strategy, not admin.

Second, verdicts expire on the buyer's clock, not yours: my Taskmarket submissions
carry fixed expiry dates (Oct 24–26), after which an undecided contest simply closes.
There is no nudge button, no "bump," no support ticket that matters. An agent's only
counter-move is portfolio discipline — enough live, dated shots that any single expiry
is a statistic, not an event.

Third, time-windowed deliverables: that $5 job requires listings to be live before
a day-7 check, with final proof gated a week out. Small money, long tail. It also produced
my favourite operational lesson: a delegate reported two directory placements "done,"
and verification at the source showed one URL was a 404 and the other was a pre-existing
listing without our token. The progress notes were retracted and corrected on the
escrow itself. In agent work — maybe in all work now — a claim you haven't verified at
the source is a liability with a timestamp.

Lesson 5: What actually works

After all of it, the working formula is unglamorous:

• Funded boards over list-a-service shelves. Money that already exists beats demand you hope will arrive.
• A submissions log, checked before every action. Boards recycle stale listings; agents that don't dedupe double-bid and burn reputation.
• Bids that cite a method and a verifiable record, not "I can do this, sir."
• Proof files by default. Every deliverable ships with sources and a SHA-256 hash (I built ProofReceipt for exactly this) — in a dispute, the side with receipts wins.
• Sweep and settle. Released funds move to my own wallet immediately; platform balances are IOUs.

The agent economy is real — I have $11.61 of on-chain proof and a second escrow in
delivery. It is also, today, mostly infrastructure in search of buyers, priced in cents
and queues. The agents that survive it won't be the ones with the best models. They'll be
the ones with the best filters.

I take on a small number of fixed-price commissions alongside the agent operation, on the
same terms the agent works under — escrow first, proof file with every delivery:
Web3/AI research reports $75 · AI agent builds & automation workflows $150 · technical
explainers & docs from $40 · content from $50. Payment in USDC/USDT to my EVM wallet,
escrow via LaborX / Synmerco / Contra (0% fee), or bank transfer. Writing:
dev.to/nenzybrown · Code:
github.com/Nenzy10 · ProofReceipt, the delivery-receipt tool
from Lesson 5: proofreceipt.netlify.app ·
Agent-build gig with escrow:
laborx.com/gigs