Research record

Research

Every dated research run behind this directory, newest first. Each report keeps the evidence, the judgement calls, the backlog and the open questions from that run, so you can see what changed and why.

What changed this week (to 7 October 2026)

New report: Agent drafts, owner submits. On Superteam Earn, payment to winners could not be verified on public routes (30 of 30 sampled); 10 agent-account wins are all unclaimed; about one agent-allowed listing appears every 11 days, so an owner can expect about $0.24 a day at any cap and about 301 days to first cash. Algora's paying boards made no new award since 2 July.
New pages and data: Is it real? Plain answers, an untested Superteam walkthrough, answers.json and human_route.json; cost translations per day and per owner hour in economics.json and the calculator.
Dated correction: TaskMarket's provisional 8.3% unpaid share (6 October) became about 21.7%; the 6 October report carries a note.
Catalogue: DeskCrew Arena added at E1 (operator's payout ledger, not independently checked); Metaculus bot tournaments not added. 21 rule rows added; 0 of 80 re-checked rows changed.
Ledger, no headline change: about $212.01 in the 30 days to 7 October (estimated; ACP roll-off not subtracted). Wallets are now shortened in the public data files.

Dated reports

7 Oct 2026Agent drafts, owner submits: Superteam payments could not be verified, and Algora's bounties stopped in JulySuperteam: 30 of 30 sampled winners wallet unknown · 10 agent-account wins, all unclaimed · about $0.24 a day at any cap, about 301 days to first cash · Algora: 0 new awards since 2 July · TaskMarket 8.3% → about 21.7% (correction) · DeskCrew added at E1 · about $212.01 to agent workers (estimated) 6 Oct 2026Funded is not paid: once your agent submits, only the requester can release the rewardTaskMarket: 23 of every 100 submissions on closed tasks went to tasks that had paid nobody when checked · 99.02 USDC locked in escrow, 0 refunded · only the requester can release a reward once anyone submits · requester check skipped 0 of 72 tasks · stranded attempts at 6 boards · $206.48 to agent workers, 29.63% operator or operator-linked 5 Oct 2026Is it worth the tokens? What one attempt is worthTaskMarket: $0.024 expected per attempt against an estimated $0.072 to $0.32 at mid prices · Superteam and Algora above cost, but pay only human accounts · win rates measurable at 3 venues · x402: 94.26% of 38,048 sellers under $1 · correction to the run 2 calculator presets · $222.84 to agent workers, 37.62% operator or operator-linked 4 Oct 2026Which agent communities are actually active, and should your agent join?17 communities measured · 6 active, 4 without a flag or caveat · verdicts: 2 join with controls, 2 for a purpose, 7 wait, 6 do not join · 1 post aimed at agents in 2,030 scanned · 6 boards re-checked, no E-level change · $223.74 to agent workers, 37.85% operator or operator-linked 3 Oct 2026Sell per call or hunt bounties? What traced buyers saySeven routes compared · 27 x402 sellers clear $1,000 from 5 buyers, 1 sells agent work (partly) · 59 x402 buyers traced: raw buyer counts overstate independent customers (correction) · 2 untested walkthroughs · $230.83 to agent workers, 40.1% operator or operator-linked 2 Oct 2026Rules, access and safety, plus ledger refresh142 quoted rules from 50 parties · 0 of 6 human-paying venues allow autonomous submission · 11 security incidents · $232.55 to agent workers, 1 identified independent payer ($9.25, promotional) 1 Oct 2026Who actually paid? Payout ledger and economics276 payouts traced · $230.37 to agent workers in 30 days · 45% from operator or operator-linked wallets · 0 identified independent payers 30 Sep 2026Where can an agent actually get paid? First dated venue catalogue33 venues · 12 with confirmed payouts · 0 with an independent payer at scale

How the research works

Who paid, not just whether. Since 1 October every payout on the agent-first boards is traced to its payer wallet and classed as operator, operator-linked, unlinked or identified independent. Since 3 October the same is done for samples of x402 buyers; raw buyer counts are never read as customers.
Evidence first. Each venue is graded on the E1 to E6 scale using primary records: public task and payout APIs, on-chain explorers, program payout records, terms pages. Secondary sources and aggregators are used for discovery only and never justify a level on their own.
Owner steps recorded step by step. Account, identity, wallet or bank, accept, submit, sign-off, withdraw.
Communities measured, not described. Since 4 October each agent community gets a dated status from its own newest items (items per day, distinct authors, top-author share), a read-only scan of those items, and a "should my agent join?" verdict from a fixed rule. Join steps are recorded from each site’s documents and stay untested.
Rules quoted and dated. Since 2 October every automation, account, KYC, maintainer and provider clause is a row in a rules matrix: quoted verbatim, given a verdict (forbid, restrict, allow, unclear), dated, and labelled by how it was read (primary, archive copy, platform JSON, or unverified snippet). Verdicts describe the text, not enforcement; none of it is legal advice.
Security from incidents, not hunches. Owner safety controls are tied to documented incidents with a primary source, plus a read-only scan of public task text.
Settlement, not just funding. Since 6 October every closed TaskMarket task gets a settlement status by a stated grace rule (awarded, final unpaid, still payable), every unawarded reward is followed into the escrow's event logs, and requester records are kept for 90 days with assumed flags. The unpaid share is always shown next to its grace sensitivity.
Per day and per owner hour. Since 7 October every computable expected value is also translated into return per $1 of tokens, attempts and expected payout per day at a stated daily cap, which constraint binds (budget or supply), median days to first cash, owner minutes per payout and payout per owner hour. Caps and owner minutes are assumed and labelled; for human-account venues the figures are for announced prizes.
Value per attempt, labelled. Since 5 October each venue gets an expected value per attempt from closed-task win rates (unawarded tasks included) with 90% ranges, and token costs estimated at three price bands. Every number is labelled measured, estimated or assumed; where a venue publishes no attempt denominator, it is marked not computable and the missing input is named.
Research only. The pool opens no accounts, takes no jobs, signs no wallets, builds no payment headers, makes no paid attempt or model call against any task and spends no money. Every pathway and walkthrough is labelled untested until a public case shows otherwise.
Corrections are dated, not silent. When a finding is overturned, the dated report keeps its original text with a visible correction note, and the change is logged.
Freshness. The space changes monthly; a status older than about 60 days is treated as stale.

Planned next

Next run (about 8 October): "Where to point your agent this week", edition 1. A graded shortlist (go with owner, probe first, avoid) across every venue with a measured verdict, each with its odds, cost translations, owner steps, requester or sponsor record and last-checked date; the grading rule is written down and back-tested before publication.
Every run: the ledger refresh, rules re-check, a light TaskMarket settlement refresh, x402 buyer refresh and the watch list. Weekly from now: Superteam overdue stock and sponsor flags, and Algora's latest award (next about 14 October). Watch: MolTrust TSK-E49N4V7T (still payable to 26 October); whether refundExpired resumes; Algora's latest payout confirmation passing 60 days on 25 October; ARC Prize Milestone #2; Steve Agent Arena winners; AgentPact stale on 31 October.
Communities: re-measure the 9 active or low-activity records around 10 to 11 October, with a light check of the other 8.
Freshness: economics rows not refreshed this run around 12 to 14 October; the oldest records (4 October) go stale at 30 days on 3 November.

Plans are provisional and depend on what each run finds.