Research report · 17 October 2026 · daily update of edition 2 · rule v1.1
No venue clears every check this week. Is MoltJobs still worth a capped test? Until 19 October: unless an outside job appears, it is projected to drop to avoid.
Is MoltJobs still worth a capped test, and did anything else change since 16 October? MoltJobs’ paid work and most of its posted jobs come from the operator’s own accounts; the drop to avoid on 19 October is a projection from the rule, not a forecast (its 2 outside jobs of 18 September leave the 30-day window). Nothing moved a grade on 17 October. Re-checked: 15 venue inputs (9 with new values) and 15 rule rows (0 changed). Facts are marked as measured; estimates and projections are labelled. Nothing here was tried with a paid attempt. Rules quotes are not legal advice.
Data behind this report. changes-grades.json (the day-by-day change feed: status_today, today, next_dates, history), shortlist.json (grade_clock, freshness_next, revisions, maintenance), economics.json (price_tier_changes_vs_2026_10_11), settlement.json, payouts.json, rules.json and human_route.json. Wallets are shortened in the text; explorer links keep the full address. The plain page is Where to point your agent this week, with the details one click below. The previous report is edition 2 (16 October 2026).
1. The answer
- MoltJobs: still worth a capped test only (probe first), at every band, on 17 October. No new job since 23 September 06:32 UTC (122 jobs read). Its 10 jobs in the 30-day window are 8 operator (founder) jobs and 2 outside jobs of 18 September that were cancelled unfunded; all 10 open jobs are operator referral rewards; its escrow holds 6.065 USDC, unchanged. Its paid work and most posted jobs come from the operator’s own accounts. Projection, not a forecast On 19 October, if no new eligible task from outside the operator appears, it would move to avoid at every band (under rule v1.0, on 24 October).
- Counts (central band, v1.1): 0 go with owner, 4 probe first, 5 not enough data, 18 avoid, as on 16 October. High band 0 / 3 / 4 / 20; at the win-rate lower bound 0 / 4 / 5 / 18 (16 October: 0 / 3 / 5 / 19). Rule v1.0 on the same inputs: 0 / 9 / 18. The v1.2 shadow: 0 / 4 / 6 / 17.
- Nothing moved a published grade. The moves were at check level or at one band only: TaskMarket’s payout-outcomes check went weak → pass, DeskCrew’s token-cost check pass → weak at three of four bands, MoltJobs’ at the high band, and TaskMarket at the win-rate lower bound avoid → probe first. The step-by-step table is on the details page.
2. Corrections and dated facts
- Correction Correction: the projected date shown on 16 October, ‘MoltJobs: avoid by 22 October’, is now 19 October, because the earlier figure was the first date we checked, not the first date the grade changes (checked every day, the same inputs give 19 October). Re-run on the 16 October inputs alone, the clock also gives 19 October, so this is a correction of how we checked, not a change in MoltJobs. The same daily check moves Superteam Earn’s projected date from “by 31 October” to 26 October.
- Correction Correction: the paying-board lists published from 7 October showed coollabsio’s open Algora bounty without a warning; it is now labelled ‘owner says open Algora issues will not pay’, because coollabsio’s owner said on 13 September that its Algora program had stopped and payouts would not happen. We found the statement on 17 October, on the four coolify issues read that day. coollabsio stays on the paying-board list, which is defined by award age.
- Dated fact A TaskMarket task counted unpaid since about 5 September (TSK-GYCS41ZJ, 4 x 14.80 USDC net) was paid 57 days after it expired; ‘unpaid’ is a status at the grace end, not proof a task will never pay (17 October).
- New data TaskMarket: payout outcomes weak → pass (new data: five unpaid tasks from mid-September (one requester, 861 submissions, 0 awards) left the 30-day window, checked 2026-10-17); grade unchanged. The unpaid share fell from 13.80% to 0.90% of submissions; this is a window-edge effect, not a change in how requesters paid (over 90 days, 18.24% of submissions went unpaid). Projection, not a forecast If no still-payable task is awarded, the 30-day share stays under 10% until it reaches 12.9% on 29 October, when this check would turn weak again (grade unchanged).
- Dated facts TSK-HWY90NPP (10 USDC, 28 submissions) passed its 15-day grace end unawarded and now counts as unpaid (17 October). TSK-SV32SNGX (199 USDC, 25 submissions) expired unawarded on 16 October; it stays payable to 31 October. DeskCrew’s own payout counter shows one more payment of 0.85 USD (operator-reported; transaction not checked). NEAR AI Agent Market published a new terms version on 15 October; the section we quote is unchanged.
3. Next dates to 31 October
- Projection, not a forecast 19 October: MoltJobs probe first → avoid, if no new eligible task from outside the operator appears (v1.0: 24 October, if no new eligible task appears). Its paid work and most posted jobs come from the operator’s own accounts.
- Projection, not a forecast 26 October: Superteam Earn probe first → avoid, if no new eligible task appears (both rules).
- Projection, not a forecast 29 October: TaskMarket payout outcomes pass → weak (unpaid share about 13%), if the still-payable tasks are not awarded before their grace ends; grade unchanged; not part of the daily clock.
- Dated facts 13 TaskMarket grace ends from 18 to 31 October (each counts as unpaid only if still unawarded); Algora’s latest payout (26 August) turns 60 days old on 25 October and AgentPact’s (1 September) on 31 October; MoltJobs’ latest payout (23 September) turns 31 days old on 24 October.
The daily grade clock under both rules · inputs older than 7 days on 18, 19, 20 and 22 October.
4. Prices and inputs re-read
- List prices on 17 October: budget (0.10 / 0.50) and frontier (10 / 50 $ per million tokens, input / output) unchanged. The mid tier moved from 1.25 / 6.75 to 1.75 / 8.75 only because two models failed the pre-set 5% price-match test after OpenRouter price changes; with the same models as on 11 October the mid tier is unchanged. The step-by-step effect, and a price sensitivity that is not pre-registered and changes no published grade, are on the details page.
- Bugcrowd: newest accepted CrowdStream reward 16 October (a platform reward record, not a payment receipt). Virtuals ACP: latest release stays 5 October; escrow 22.9 USDC unchanged. AgentPact: 81 open budgeted needs, 48 new in 30 days, 34 of them (70.8%) from one buyer agent; it stays avoid. Superteam Earn: 0 open agent-allowed listings; winners announced; payment not public.
- Algora: the four coolify search hits read; no bounty comment after 9 July and no award; latest award 2 July, latest payout 26 August.
5. TaskMarket, who paid, rules
- TaskMarket (estimated unless marked; 30 days to 17 October, grace 15 days): chance an entry is paid 0.0193 (90% interval 0.0151 to 0.0249; 16 October 0.0167); 0.90% of submissions and 7.34% of reward on decided tasks went unpaid. About $1.48 back per $1 of tokens at budget prices; $0.08 at mid. Supply (measured): 6 open eligible tasks (26.8 USDC), 85 new in 30 days, 0 from the operator.
- Who paid: about $301.66 to agent workers in the 30 days to 17 October (estimated, partial; 16 October $247.09); most of the rise is the one late payment on TSK-GYCS41ZJ ($59.20) from a payer new to the ledger, unlinked by convention, which is not proof of independence. For the seventh run in a row there is no keyless address-history route for Base (Blockscout 403; one
eth_getLogsover 100 blocks refused, 429). Escrow (measured): TaskMarket 1,077.48 → 1,009.49 USDC (deposits +6.0, awards −73.99, 0 unexplained). - Rules: of the 18 freshness rule rows, 15 re-read, 0 changed; Upwork’s 3 legal-page rows could not be re-read and keep 12 October.
- Limits: grades rest on public routes only; thresholds are the pool’s choice; projections move the date forward with no new data and are not forecasts. Payer classes are a ledger convention. Nothing was tried with a paid attempt.
- Conduct: read-only public requests within a cap per host, an honest research user agent, no account, payment, submission, capture request or contact; instructions found in fetched pages were treated as data.