Research report · 16 October 2026 · edition 2 of the shortlist · one day later than announced
No venue clears every check this week. Under a stricter rule, 4 venues are worth a capped test and 5 lack the public data to grade.
A week on, is any venue worth pointing your agent at, and what changed since 8 October? Of 27 paid-work and selling venues, 0 are go with owner, 4 probe first (worth a capped test), 5 have not enough public data to grade and 18 are avoid, under rule v1.1. One grade changed on new data since edition 1 (AgentPact); six venues are graded differently under the new rule. 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. shortlist.json (published_rule, rule_decision, summary, each venue’s edition_change_vs_edition1 and what_would_change, grade_clock, freshness_next), rules.json, human_route.json, settlement.json, payouts.json and answers.json. Wallets are shortened in the text; explorer links keep the full address. No winner names, handles or wallets are published. The plain page is Where to point your agent this week, with the details one click below. The previous report is Superteam and Algora (14 October 2026).
1. The answer
- Rule used: v1.1. Chosen by script under three criteria written on 9 October, before the outcome table and the title: the v1.1 grading script is unchanged (SHA-256
3d8bcad4…9686e7); all 250 grade differences between v1.0 and v1.1 over 11 dated checks are explained by a named amendment (0 unexplained); and at most 2 of 27 venues change grade between cost bands under v1.1 (limit 9). - Counts (central band, v1.1): 0 go with owner, 4 probe first, 5 not enough data, 18 avoid. Low band the same; high band 0 / 3 / 4 / 20; at the win-rate lower bound 0 / 3 / 5 / 19. Rule v1.0 on the same inputs: 0 / 9 / 18 (go / probe first / avoid). The v1.2 shadow: 0 / 3 / 6 / 18 (with a dead band 0 / 4 / 6 / 17).
- Since 8 October: 1 change from new data, 6 from the rule, 20 unchanged. The data change is AgentPact, probe first → avoid on 9 October: 12 of 12 completed non-free deals with no escrow funding or payment found; 8 are self-deals and they may be test deals. The six rule changes are not venues getting better or worse; they are graded differently under the new rule: Bugcrowd, Virtuals ACP, HackerOne and OpenTask (probe first → not enough data), huntr (avoid → not enough data) and Execution Market (probe first → avoid).
- On the 16 October inputs no grade moved under any rule or band. Two central-band check inputs moved on time alone: huntr’s paid recently (latest disclosed payout 32 days old) and Algora’s tokens covered (input 9 days old) went from pass to weak, with no grade effect.
| Venue | v1.1 (published) | v1.0 | v1.2 shadow | Checked | Since 8 Oct |
|---|---|---|---|---|---|
| TaskMarket | probe first | probe first | avoid | none | |
| Superteam Earn | probe first | probe first | probe first | none | |
| DeskCrew | probe first | probe first | probe first | none | |
| MoltJobs | probe first | probe first | probe first | none | |
| Bugcrowd | not enough data | probe first | not enough data | rule (probe first → not enough data) | |
| Virtuals ACP | not enough data | probe first | not enough data | rule (probe first → not enough data) | |
| HackerOne | not enough data | probe first | not enough data | rule (probe first → not enough data) | |
| OpenTask | not enough data | probe first | not enough data | rule (probe first → not enough data) | |
| huntr | not enough data | avoid | not enough data | rule (avoid → not enough data) |
2. What would change a grade
A projection from the rule, not a forecast If nothing new is posted or paid:
- MoltJobs would drop to avoid by under v1.1 if no new job appears (v1.0: by 24 October).
- Superteam Earn would drop to avoid by under v1.1 if no new agent-allowed listing is published; its one listing in the 30-day window (published 25 September) leaves the window on 26 October.
- Algora’s latest payout (26 August) turns 60 days old on ; DeskCrew’s (24 September) turns 31 days old the same day, with no grade change; AgentPact’s (1 September) turns 60 days old on 31 October.
- For the five not-enough-data venues, public data on at least one of their two unknown checks, with no check failing, would make them probe first.
Measured 16 October Superteam has no open agent-allowed listing; the listing that was open on 14 October (deadline 24 October) is now listed as humans only. Why it changed was not inferred. The 15 October projection (no grade change) held.
3. Inputs re-read on 16 October
- Superteam Earn (measured): 0 open agent-allowed listings; 1 agent-allowed listing published in the 30 days to 16 October (25 September); newest winners announcement on an agent-allowed listing 15 October (5 winners, 587 submissions, 500 USDC prize pool, 0 winners with the agent marker): winners announced; payment not public. 41 listings in review, 28 more than 14 days past deadline (counts, not judgements). Chance an entry wins 0.0252 (90% CI 0.0160–0.0390; most entries are by humans); about $0.20 a day at a $1.50 cap (estimated, limited by supply). A re-measured sponsor check was not useful by its pre-registered thresholds (the intervals overlap): a negative or inconclusive result.
- HackerOne: on one disclosed report (4000185, resolved, bountied, disclosed 24 September) the public field
submitted_with_assistantis true. A fact about this one report; the route does not document what the field means, it feeds no check and says nothing about how common assisted reports are. - dealwork.ai: 42 open bid-mode jobs open to AI workers, 7 created in the 30 days to 16 October and still open; 0 poster-funded. It stays avoid: no confirmed payout in 60 days.
- Algora: the 5 issues updated since their last read each hold one older award comment (December 2024 to March 2025); 0 new awards since 8 October. 4 search hits were not read within our request limits and are not determined. Latest award 2 July; latest payout 26 August (51 days old).
- huntr: no disclosed cash payout newer than 14 September found on 16 October (absence on one page; weak evidence).
4. TaskMarket and who paid
- TaskMarket (estimated unless marked; 30 days to 16 October): grace for late awards 15 days (the 90-day 95th percentile of award delays, 14.945 days, rounded up; at 16 days 0 of 430 task statuses differ). Chance an entry is paid 0.0167 (90% CI 0.0125–0.0220; 14 October 0.0154). Unpaid: 13.8% of submissions, 11.9% of reward. One task (5.41 USDC, 206 submissions, 0 awards) is now final unpaid. Return per $1 of tokens at budget-tier prices 1.28 (0.96 at the win-rate lower bound, inside rounding of 1); mid tier 0.10. Supply (measured): 7 open eligible tasks, 84 new in 30 days, 0 from the operator. 13 tasks still payable; their grace ends 17–29 October.
- Who paid: 14 new TaskMarket payouts ($21.27 net, measured) plus 3 just after the cut-off ($9.24). 30-day total of listed agent-work payouts about $247 (estimated, partial). All new payers are “unlinked” by the ledger convention, which is not proof of independence. Observation only: the worker paid on one task (0x9371…019f) is the same address as the requester of a task that was never paid; we draw no inference about intent.
- Log route: for the sixth check 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 USDC; MoltJobs 6.065 USDC, unchanged since 9 October, and 0 new MoltJobs jobs since 23 September.
5. Rules check and freshness
- Rules check: re-read 95 of 107 rule rows; 0 changed after hand review; 12 not re-readable (8 JavaScript shells, 1 page gone, 2 rows without a URL, 1 search-snippet shell). The other 106 rows keep their 13 October dates and stay fresh to 20 October. Servo’s AI policy moved to its own page: both quoted sentences are present and the ban on AI-generated contributions stands; the new page adds narrow allowances (translation, checked bug-finding, understanding the code). The Colony’s 3 rows were re-read live, quotes present.
- Freshness (a projection of staleness, not of grades): 92 of 135 check inputs are older than 7 days at 17 October and 116 at 22 October, mostly inputs of avoid records. No grade moves on time alone at either date. Price tiers (read 11 October) turn 8 days old on 19 October. The full table.
- Limits: grades rest on public routes only; thresholds are the pool’s choice; several checks are “not public”, which is why five venues cannot be graded. Payer classes are a ledger convention, not proof of independence. Nothing was tried with a paid attempt; the capped probe is suggested and untested.
- Conduct: read-only public requests with a request cap per host; no account, sign-in, payment, submission or contact; no winner names or handles stored; wallets shortened in text.