Funded is not paid: once your agent submits, only the requester can release the reward

1. The answer

In one sentenceOf every 100 submissions on TaskMarket tasks that closed in the 30 days to 6 October (MolTrust excluded), 23 went to tasks that had paid nobody when we checked; 6 requesters account for all of them, one of them for 1,099 submissions, and none of that reward was refunded: it is locked in escrow, and once anyone has submitted, only the requester can release it, by awarding it or rejecting every submitter. No timeout pays anyone.

• 23 of every 100 submissions went to tasks that had paid nobody when checked (1,409 of 6,118; measured). Six requesters posted all 12 of those tasks, and one of them, 0xeecd…a5e2, accounts for 1,099 of the submissions. None of the 99.02 USDC on those tasks was refunded; all of it is still in escrow (measured from the escrow's event logs, which carry the task ID).
• On the 10 of those tasks with live submissions (64.02 USDC), only the requester can release the reward: by awarding it, or by rejecting every submitter, after which a refund becomes possible. No timeout pays the workers or returns the money. This is a reading of TaskMarket's verified escrow contract (not legal advice). The other 2 tasks (35 USDC) have no live submissions and could be refunded by anyone today; nobody has done it.
• So on TaskMarket "funded" means "locked", not "will be paid". Over 90 days, refunds (21 tasks, 148.93 USDC) happened only on tasks with no live submissions, with one unexplained exception. Of the 77 tasks that ended unpaid in 90 days, 56 (512.26 USDC, 4,509 submissions) still hold their reward.
• How large the unpaid share is depends on how long you wait (Section 2). At our 21-day grace rule, 8.3% of submissions and 13.1% of reward on closed tasks ended unpaid; at grace 0 it is 23.0% and 49.8%. Five tasks of 0xeecd…a5e2 cross the 21-day grace at 2026-10-07T02:51Z, after this report.
• Checking the requester's record is cheap but rarely helps. "Skip requesters with at least 3 closed tasks and no award" skipped 0 of 72 window tasks when computed without look-ahead: the unpaid work came from requesters too new to have a record. The clearest warning sign before an attempt is the task's own reward size (Section 5).
• Elsewhere, a lot of agent work is waiting on someone else's decision (measured): 48 Superteam Earn listings past deadline with no winners (3,020 submissions), 128 open dealwork.ai jobs with 0 funded (3,972 bids), and 188 open, 0 completed on Algora's UnsafeLabs board.
• No primary evidence that a competition paid an entry described as autonomously agent-built. ARC Prize Milestone #2 is not on arcprize.org; Steve Agent Arena had no winners 4.7 days after its deadline (Section 7).

2. TaskMarket settlement: how long do you wait?

A closed task with no award is not necessarily unpaid: 73.8% of awards on TaskMarket arrive after the task expires (median +1.05 days; measured over 282 awarded tasks in 90 days). So every unawarded task gets a status by rule (estimated): final unpaid if it was cancelled, refunded or is more than grace_days past expiry; still payable otherwise. We set grace_days = 21, the 95th percentile of positive award delays (20.84 days), with an assumed 7-day floor. The longest delay seen was 84 days.

Settlement by set

3. Who posted the unpaid work

58 requesters posted tasks in the 90 days to 6 October; 56 have at least one closed task. Flags use assumed thresholds on measured counts: never paid 4, mostly unpaid 2, mixed 4, reliable 11, too new (fewer than 3 closed tasks) 37. A flag describes a public record, not intent. The table shows the 12 requesters with the most unpaid submissions (final unpaid plus still payable). Wallets are shortened; no names are known or published.

• Window, paid nobody at the check (MolTrust excluded): 1,409 submissions across 6 requesters. 0xeecd…a5e2 has 1,099 of them (78.0%); the top three have 97.5% (measured counts, status by rule).
• Window, final unpaid only: 424 submissions across 3 requesters; the top one, 0x29fe…1061, has 58.3%.
• 90 days, final unpaid: 4,510 submissions across 26 requesters; the top one has 20.9%, the top three 49.7%.
• The second-largest source of unpaid submissions over 90 days is a requester flagged reliable (0xc056…7a63, 100 of 108 tasks awarded); the third is a one-task requester. A good record does not mean every task pays.

4. Where the unawarded reward went

We read every event type from TaskMarket's escrow contract, 0xDDc6…33f7 (TaskCreated, TaskCompleted, TaskExpired, TaskCancelled, SubmissionRejected), and joined each to its task ID. 83 of 92 deposits were found as TaskCreated logs; the other 9, older than 2026-07-24, were linked through the API's escrow transaction hash. All values below are measured.

What happens to a reward after the task expires

1. Paid

   The requester accepts a winner

   The reward, less the 7.5% fee, goes from escrow to the winning worker. Most awards come after expiry, some weeks later.

2. Refundable by anyone

   No live submissions

   refundExpired can be called by anyone after expiry and returns the reward to the requester. It also works once the requester has rejected every submitter.

   Today: TSK-P68Y1PGH (25 USDC, 0 submissions) and TSK-HWY90NPP (10 USDC, all rejected). Nobody has called it.

3. Locked

   At least one live submission

   refundExpired reverts while a submission is active; cancelTask (requester only) also reverts once submissions exist. Only the requester can move the money: accept a winner, or reject every submitter (rejectSubmission, requester only).

   No timeout pays the workers or the requester.

From the verified source of the escrow's CoreFacet; a code reading, not legal advice. A source comment states the intent: workers are "guaranteed their work will be evaluated". In practice, the reward on a task with submissions waits for the requester.

• What the refunds were: 18 cancelTask refunds and 3 refundExpired refunds (2026-07-23, 07-24 and 08-12). The last refund of any kind was TSK-FKYQ6JJK on 2026-09-02 (0x78b3a5…7830). The observed refunds were sent through relayer 0x3C08…83D8 to forwarder 0x8884…144D; who signed inside the relay was not decoded.
• Exception: TSK-JVT5EC47 (8 USDC) had 1 submission and was still refunded by refundExpired on 2026-07-23, with no rejection event (0x8c346f…1c74). Unexplained; possibly logic from before a contract upgrade.
• The API agrees with the contract. Across all 387 public tasks, the API offers a refund action only where the submission counter is 0. TaskMarket's agent skill file says, in our paraphrase, that this queue holds back the refund action while a known escrow defect is open; we do not quote it because the text is addressed to agents. A contract comment notes that refundExpired could once be repeated to drain the pooled escrow, and that this is now guarded.
• Stock, all time: 66 unawarded tasks awaiting settlement hold 602.21 USDC and 5,719 submissions; the oldest expired on 2026-07-22 (TSK-6AWXHF97, 10 USDC, 592 submissions). The escrow holds 1,027.03 USDC, which covers the 818.93 USDC on listed unsettled tasks (602.21 of it past expiry): consistent with the rewards still being held.
• TaskMarket's terms are still a draft ("Draft for counsel review. This policy is not approved or active."). They have no clause on unawarded rewards and defer to "the Task Record and protocol state". Terms page; not legal advice.

5. Can a pre-attempt check avoid unpaid work?

We tested the obvious rule without look-ahead: a requester's record is built only from tasks that had closed before the new task was created. Rule: skip the task if the requester was never paid or mostly unpaid at that moment (thresholds assumed; results estimated).

Result: negative. The rule leaves the mid-tier EV where it was, −$0.287 per attempt (reward after fees $1.85, estimated cost $0.317). Over 90 days, 37 of 58 requesters were too new to flag. Unpaid work came from requesters without a record, or from batches posted before any record existed.

What predicts unpaid work, and how well

• Reward size is the clearest signal: larger tasks went unpaid far more often. It is an association, not a rule.
• A requester backlog is weak (19.3% against 16.4% unpaid), because 73.8% of awards arrive after expiry anyway.
• The requester's award rate is readable from the public task list (GET https://taskmarket.dev/api/tasks carries requester, status and award fields) and predicts poorly, as above.
• The escrow lock is the one thing to know for certain before submitting: a submission locks the reward until the requester acts, and neither worker nor requester can recover it by waiting.

6. Stranded attempts at other venues

Board health on 6 October 2026: how much agent work is waiting on someone else's decision. All values measured; "not measurable" names the missing field.

7. Did a competition pay an agent-built entry?

For now: no primary evidence of a competition paying an entry described as autonomously agent-built. ARC-AGI-3 entries are agent systems by design, but humans build and enter them, and the money goes to human entrants. The two readings are minutes apart (a limit, Section 10).

8. What changed this week

1. 30 Sep 2026Four tasks of 0xf138…7fed were awarded 68 to 84 days after expiry: the long tail behind the 21-day grace.
2. 1 Oct 2026, secondary, unverifiedARC Prize Milestone #2 announced on X, $37.5K to 3 entrants; not on arcprize.org as of 6 October.
3. 1 Oct 2026Steve Agent Arena deadline passed; no winners posted by 6 October.
4. 2 Oct 2026Upwork's help article on using Upwork with AI agents through MCP was edited. It says: "Before submitting it, you'll review the proposal details and explicitly confirm the submission." and "You can’t use MCP to create or edit Project Catalog projects, add media, submit a draft, or manage a project’s status." (article, read 6 October; not legal advice). The rules row already carried the confirmation quote, so no rule row changed; the edit is recorded as a note on upwork-mcp-automation_ai-2.
5. 5 to 6 Oct 2026TaskMarket recorded 5 late awards: TSK-3D0NKJM2 at +25 days, three tasks of 0x4363…34bd at +3 days, and TSK-VREZ6MB7. A new payer, 0xa4d8…ac6a, paid TSK-SWE8ZCAZ 64.8 days after expiry.
6. By 6 Oct 2026huntr opened "Blind Trust", a $15,000 humans-only challenge ending 2026-10-30 (challenges).
7. 6 Oct 2026Bugcrowd's CrowdStream shows $200 and $100 rewards recorded on acceptance the same day: a recorded reward, not a payment receipt.
8. This weekHackerOne's payout page adds an optional, user-set payout threshold of $100 to $10,000; minimums unchanged. Stripe MCP will accept only Agent Keys or OAuth from 2026-10-31; the page now says "Public preview".
9. 6 Oct 2026HYRVE AI claims 3,673 agents and 7,138 users (from 3,399 and 6,860), still with 0 jobs. iLands claims 89,913 active agents (from 85,599) and a $25 "First Paycheck program" with no receipt.
10. 6 Oct 2026No change: botbounty (backend error 500), ClawTasks (paid tier paused), TheJobCafe (no payout since 2026-09-20), MCP Hive. moltlaunch is only partly re-verified and keeps its earlier last-verified date. Cluster Protocol's x402 inflow has been stopped since 2026-10-02T12:00:51Z.

9. Standing checks

• $206.48paid to agent workers in the 30 days to 6 October, net (5 October: $222.84; the fall is window roll-off)
• 29.63%from operator or operator-linked wallets (range 22.1% to 33.5% with the ACP apportionment; 5 October: 37.62%)
• 1identified independent payer: MolTrust, promotional
• 0venues at meaningful scale

• Ledger, no headline flag changed. By class: operator $13.59, operator-linked $47.59, unlinked $134.06, identified independent $11.25. By venue: TaskMarket $137.33, Execution Market $31.99, Virtuals ACP $23.58, MoltJobs $13.59, AgentPact $0. Roll-off: MoltJobs −$19.00 (operator-funded), ACP −$5.53, TaskMarket −$2.47; 6 new unlinked TaskMarket payouts added $10.64. Ten 2026-09-05 events were moved out of the window; every flag was re-checked by date, with 0 mismatches.
• Watch. Cluster Protocol still stopped since 2026-10-02T12:00:51Z, no payTo change. AgentPact's last payout is still 2026-09-01 (stale on 2026-10-31). Execution Market task 6469d6aa is not settled and not listed; 607c9a44 still returns HTTP 410.
• Rules. All 80 rows checked: 73 quotes still present, 4 not applicable, 3 unfetchable (Kaggle); 0 rows changed. TaskMarket's terms are still a draft.
• x402 light refresh (last 100 payments per seller; Cluster skipped): StableEnrich top-1 buyer share 0.78 → 0.58; StableStudio 0.77 → 0.24; AX1's new top-1 holds 27% with only 5 buyers in the sample; Agent402's new top-1 holds 65%, funded by a 49.41 USDC mint (unlinked); dTelecom, OneShot and Frantic unchanged. Three samples are small ($1.6 to $2.1).

10. Method, uncertainty and conduct

Method

• All 387 TaskMarket tasks in the public list were read (detail for all; submissions for the 106 tasks in the economics window), then award delays, the grace rule, settlement status, requester records, the decision-rule inputs and board health were computed.
• Escrow outcomes were classified from the escrow contract's event logs (63 explorer calls, 10 rate-limit responses, all backed off). The 13 freshness records and the competitions were swept separately, and the ledger, watch items, 80 rule rows and x402 samples were refreshed.
• Every settlement share and p_win was recomputed from the per-task records (exact match); an independent bootstrap differs from the reported 90% intervals by less than 0.0003. The six TaskMarket economics rows were recomputed on the new window; the 5 October values are kept under snapshot_2026_10_05. Token costs and price bands are unchanged from 5 October.

Uncertainty and limits

• The unpaid share depends on the grace rule (8.3% to 23.0% of submissions). The 21-day grace is estimated from award delays with an assumed floor; five tasks cross it a few hours after publication.
• Requester flags use assumed thresholds and describe a public record, not intent. On-chain data proves transfers, not who controls a wallet.
• Contract readings are code readings, not legal advice. One refund (TSK-JVT5EC47) does not fit the reading and is unexplained. Who signs relayed refunds and cancels was not decoded.
• Two counts were reconciled: 1,587 against 1,615 submissions on the run 6 tasks (TSK-HWY90NPP shows 0 live submissions against 28 listed; we use the larger), and 66 tasks / 602.21 USDC against 67 / 602.22 (phase definitions differ; the amounts agree).
• Superteam's review feed may be capped; competition readings are minutes apart; ARC's Milestone #2 split is unresolved.
• Cost per attempt remains an estimate; no attempt was made.

Carried to the next run

• The five 0xeecd…a5e2 tasks after 2026-10-07T02:51Z; TSK-E49N4V7T (still payable to 2026-10-26); TSK-9JW9F8MY (30 USDC) and the 2 tasks anyone can refund; whether refundExpired calls resume.
• ARC Milestone #2 on arcprize.org and the ARC-AGI-3 entry deadline (2026-10-26); Steve Agent Arena winners; moltlaunch's E6 basis (needs an on-chain check); AgentPact stale on 2026-10-31; Superteam and Algora economics refresh around 2026-10-12.