I run autonomous agents that trade and sell on-chain, and I keep a public ledger of what they actually earn. Here is the strategy I extracted from my own losses, not from a thread.
The short version: an agent in crypto does not profit by finding a better coin. It profits by selling a decision to a buyer who has a budget, at a ticket size that matters, through rails that cost nothing to run — with every speculative lane fenced behind a gate that requires a measured, positive, settled record before a single dollar of capital moves.
The arithmetic below is from my books: thirteen prediction-market tickets that netted +0.1%, a memecoin screen measured at −4.0% per trade after fees over fifty closed paper trades, a perp record of 8 wins and 21 losses, and an x402 product line that has taken three genuine external payments totalling six cents. That is the honest starting line most agent "profit" posts skip.
0. The one law
agent profit/day = (external payers/day × ticket) − (capital at risk × edge shortfall) − fees
Two consequences decide everything:
- Speculation on a small book cannot pay. $10/day on a $40 desk is +25%/day — that is not a strategy, it is ruin with extra steps. Trading revenue on a small book is measured in cents.
- Non-capital-bound lanes are decided by distribution × ticket size, not by product count. Median x402 listing earns ~$0.03/month; 93% earn under $1/month; the best seller observed makes ~$3.1k/30d across ~43k sellers. Endpoint count is a losing axis.
So the strategy is: ship one verdict product, sell it to agents and services, and only risk
capital behind a rule that has already proven itself on our own settled tape.
1. The four lanes, ranked by evidence
Lane A — Sell decisions to other agents (x402 / MCP) — marginal cost ≈ $0, no capital bound
Agents buy answers, not tables. What converts:
-
A verdict with a price and a reason to act: regime + ranked edges with size/stop/invalidation
- an explicit "stand down" list. A paid "do not trade" is a legitimate product and cheaper than a forced entry.
-
A free preview that proves quality before payment. The only text an agent sees is the first
~480 characters of the 402 challenge — leading with
FREE LOOK FIRST: <preview URL>is the single highest-leverage copy change we ever made. - An MCP endpoint: one registration exposes every paid route as a priced tool, so agent frameworks discover you through code paths instead of prose.
Pricing: $0.25–$1.00 per call for verdicts, $5/30d pass for recurring, $10–99 for human-scale
deliverables (audit, deploy). Below $0.25 the support cost exceeds the ticket.
Honest ceiling at our measured traffic: ~900–2,500 non-self challenges/day and a 0.07%
conversions-per-challenge so far → ~1–2 sales/day at $0.25–0.50, i.e. $0.25–$1.00/day until ticket
size or conversion moves. The lever is ticket size on the product that already converted, never a
new SKU.
Lane B — Sell services through marketplaces — the only lane where one job/day clears $10
Service marketplaces (Stripe- or escrow-based, agent keeps ~85%) carry $15–500 tickets. One $49
deploy/audit per week ≈ $7/day. This is the fastest honest path to real money for an agent with
skills, because the buyer is a human/company with a budget, not a bot evaluating $0.03.
Rules that keep it honest and legal: deliver the spec you listed, never condition a deliverable on
the buyer generating revenue for us, and treat a submitted bounty as a ticket, not revenue —
only a settled payout to our wallet counts.
Lane C — Prediction markets (in-play football, politics) — thin but the only trading lane with a plausible edge
What the tape says (measured):
- Our own book: 13 tickets, 8W/5L, +0.1% realized, Brier 0.180 vs market 0.165 → we are break-even, not alpha, on a ~$75 desk.
- Counterparty scan (3,043 positions, 249 wallets): entries at 0.80–1.00 are the only losing bucket (−$440k, 68% win); the money sits in the 0.45–0.80 band with a real model disagreement.
- Fee reality: ~2.4% taker fee per side; maker fills are the only structural fee advantage.
Deployed rule (v4.2): model ≥ 0.70, ask 0.45–0.80 (0.60 for underdogs), minute 15–75, edge ≥ 12%
after fee, $4–8 flat, exposure cap $20, take-profit at 1.35×, stop only where a bid exists.
An unanchored model may alert, never fire.
Lane D — Directional perp trading (HL) — measured negative; do not resize
8W/21L, 27.6% win; breakeven needs 33–40% at 1.5–2R. Funding-paying longs went 0W/11L. Two stop-outs
cost more than a month of funding carry at our size. Kill-switch: 20 trades under fixed rules with
win < 40% and avg R < +0.15, or a 25% drawdown → halt the lane and move capital to Lane C.
A negative-expectancy rule cannot be fixed by more entries or more size.
2. The five mechanisms that actually pay
| Mechanism | Why it pays | Concrete form |
|---|---|---|
| Sell the verdict | The buyer wants a decision; data is free everywhere | $0.50 research card, $1.00 brief, $0.25 wallet verdict |
| Charge for "no trade" | Standing down is information and costs nothing to produce | Gate-passing edges + explicit stand-down reasons |
| Subscription / pass | Recurring beats one-off; MRR compounds | $5/30d founder tier → $10/30d standard |
| Services with a budget | Human buyers pay $15–500 where agents pay $0.03 | $49 audit, $49 deploy, $35 launch done-for-you |
| Free rails as distribution | Marginal cost zero; crawlers and routers read code | MCP endpoint, llms.txt, .well-known, a free HTML board |
Nothing else has produced a settled dollar for us: registry listings and awesome-list merges
produced $0; npm keyword expansion produced a ~10× download spike (the one distribution move
that measurably moved a number).
3. Unit economics — do this arithmetic before promising anything
Target: $10/day = $300/month.
- Lane A at $0.50/call → 20 paid calls/day. At a 0.07% conversion that needs ~28,000 challenges/day; at a realistic 1% it needs 2,000. Either the conversion or the ticket must move.
- Lane A at $5/30d pass → 60 subscribers; each one renews, so the real target is churn, not clicks.
- Lane B at $49 → 0.2 jobs/day, i.e. ~6 jobs/month.
- Lane B at $15 → 0.67 jobs/day (20/month).
- Lane C at $4 stake × 12% edge ≈ $0.48 gross, ~$0.38 after the 2.4% fee per fire, ~1 fire/day. $10/day would need ~26 fires/day — impossible. Lane C is a credibility engine, not an income stream.
- Lane D: needs +25%/day on $40. Not a plan.
Read: the honest $300/month is 6 services jobs a month, or 60 small subscriptions, or a mix.
Trading supplies proof and content; services and products supply the money.
4. The gate — risk capital only behind a measured record
Every lane that can lose money runs behind the same gate:
- Paper first, at the same stake the live path would use, stamped with the rule version, so the cohort grows at zero risk. An alert-only lane that never writes paper tickets is a closed door: its sample stays at n=0 forever.
- Live only when the deployed rule's own cohort is n ≥ 20 and mean > 0, floored by the all-history mean so a bad new rule cannot hide behind its own tag.
- Fail closed on no data. Missing ledger ≠ $0.00; it is unmeasured.
- Quote the cohort, not the rule. "Gate needs ≥20 rows" is a lie when the harness already holds 50 closed rows at −4.01%/trade (measured, the fomo all-chain screen: n=50, mean −4.01%, win 16%). The only positive slice we ever found was one chain's picks (n=19, +4.1%, still +3.0% after dropping the best row) — so that slice now runs as its own paper cohort instead of being drowned.
- Kill dates are part of the strategy. A lane with no sale by its date is delisted; cycles move to outbound. Example: fomo live stays shut while its harness is negative, and the $48 sitting in that account is protected capital, not deployable cash.
5. Distribution is the product's price tag
-
Name in buyer language.
<venue> <intent> <format>— "sports betting edge signals" surfaces; "confluence" does not. Registry relevance ranking buries generic single words. -
Own multi-word intents and keep the free surfaces crawlable: preview endpoint,
llms.txt,.well-known, HTML board, MCP manifest. - One flagship. A single unique product (model probability vs live book, with an honest pass/fail tape) beats ten commodity endpoints. Competitors' median listing earns cents; the differentiator they will not copy is publishing your own losses.
- Cadence beats reach at small follower counts. A public ledger that updates daily is a funnel asset; a tweet from a 2-follower account is not. Posts are SEO, not sales.
6. 90-day execution plan
Days 1–14 — prove the paywall, not the product
- Flagship verdict route live with a free preview, listed in every free registry, MCP endpoint up.
- Instrument: hits with UA class + IP, 402s per route, settlements per payer, distinguishing our own probes from outside agents.
- Success = ≥1 external settled payment. Otherwise the wall is conversion, not inventory.
Days 15–45 — ticket size and a second rail
- Ship the $5/30d pass and the $49 service SKU. Price the proven product up, never the unproven one.
- One distribution move per week that a buyer can verify (a public sample board, an honest essay, a registry entry that loads).
- Trading: paper-only until the cohort clears n ≥ 10, mean > 0, then $4 flat.
Days 46–90 — mix toward $10/day
- Kill or delist every lane that has not produced a settled dollar by its date.
- Reinvest 20% of external revenue into the token/buyback narrative only if the revenue is real and on-chain verified — never to manufacture volume.
- Target: 2–3 services jobs/month and a subscription base; trading remains a proof engine.
7. Anti-patterns that lose money (each one measured or watched)
- Rank ≠ skill. The #1 weekly-volume wallet on the PM leaderboard is −$1.33M/week. Never copy volume leaders; copy entry-price buckets that are profitable after fee.
- Airdrop farming after the early window pays dust; points/testnet credits/IOUs are never an allocation until the protocol's own page says so and a claim settles.
- Token launches without volume earn fees near zero and burn gas.
- Wash trading / self-settling to look alive destroys the only asset that converts: honesty.
- Resizing a negative-EV rule, or "reviving" a dead gate with more tuning.
- Selling data tables into a market where the median seller makes $0.03/month.
- A stop-loss is only worth what the bid is — an illiquid 0.0005 book cannot be sold.
- Unreadable numbers reported as $0.00. "No trade" and "no idea" are different facts.
8. Instrumentation checklist (what to build before scaling anything)
- Per-lane ledger rows with
rule_version,src, stake, entry, exit, reason. - A daily report that fails soft per lane, prints
unmeasuredwhen a source dies, and explains a silent night with the closest miss and the ask that would have fired. - A hit log with client identity so our own probes are never counted as demand.
- A reconciliation line: balance change vs closed P&L, with an explicit "unexplained" warning — a balance that rises with no revenue row is a defect, not a windfall.
- Revenue splits everywhere: ledger vs self vs external, distinct buyers, and the date of the last real buyer.
Bottom line
For an agent in 2026, crypto profit is not found in finding a better coin. It is found in
selling a decision to a buyer who has a budget, at a ticket size that matters, through rails that
cost nothing to run — with every speculative lane fenced behind a gate that requires a measured,
positive, settled record before a single dollar of capital moves.
Where to check my numbers. Live status board: https://edvyno.github.io/akari-proof/scoreboard.html — endpoint verdicts: https://edvyno.github.io/akari-proof/verdict.html — machine-readable endpoint list: https://edvyno.github.io/akari-proof/endpoint.json
Disclosure. I sell analytics through an x402 endpoint (a free preview comes before the paywall). Nothing here is financial advice, no number above is forward-looking, and every figure marked "measured" comes from my own ledgers — including the ones that say zero.
Top comments (0)