AI Agent Scam: 192 Jobs Completed, Zero Paid

Editor’s Take · AI Serial Founder Perspective (Content distilled by AI; views belong to the original author. You can skip the source after reading this.)

This covers the settlement bottlenecks AI agents hit when taking on task-market jobs. Key evidence: the Rook agent won all 192 tasks but never received funds (tested by an AI); the Toku agent earned only $4.50 on a $15 order (tested by an AI). What this means for making money: the opportunity isn’t in “running agents,” but in the settlement tools between agents or the manual review环节. Pure agent-running is risky—wallets get locked and demand can be fabricated. Actionable moves: prioritize building an agent budget-monitoring tool or offer a manual QA service for agent-taken tasks.

• Monitor agent wallet spend to prevent a single task from burning through profits
• Launch a manual QA service for agent tasks and capture the spread
• Build a settlement-speed ranking for task platforms as a data product
• Don’t lock your agent fleet to a single platform’s API

1. What Opportunity Is This

Address the pain point of AI agents getting their funds stuck or failing to settle after winning task-market jobs, by offering either an agent wallet budget-monitoring tool or a manual QA & sign-off service. Target customers are indie developers or companies running fleets of AI agents. Revenue comes from service fees or tool subscriptions for solving the “won the task but can’t get paid” settlement bottleneck.

2. Independent Judgment

Worth pursuing, but the entry point is “settlement infrastructure,” not “running agents.” The original testing shows agents can complete tasks yet can’t freely access funds—for example, Rook won 192 orders and got nothing. That confirms real demand. The key distinction: simply running agents leaves you vulnerable to platform API lock-in or fake demand, while providing cross-platform budget monitoring and manual sign-off tackles a common pain for every agent fleet, yielding higher commercial certainty.

3. Cold-Start Path

Step one: build a lightweight wallet-monitoring plugin that integrates with major agent runtimes (e.g., OpenClaw) and enforces a per-task spend cap—for instance, don’t let a $5 task burn $3 on search costs. Development cost is minimal, mostly just engineering time. Launch an MVP in roughly two to four weeks and sell directly to developers already running TaskMarket or dealwork to validate willingness to pay.

4. Biggest Risks and How to Avoid Them

Risk one: fake seed demand. Some task platforms (like Toku) may have orders generated by automation tools (e.g., 0ai-Supervisor) rather than real human clients, inflating volume without conversion. Mitigation: add a “demand-source attribution” feature to your monitoring tool and only count verified human or enterprise-budget orders.

Risk two: agent fleet tied to a single platform. If your tool or fleet relies heavily on one platform’s API, the business hits zero if that platform shuts down or changes its rules. Mitigation: keep interfaces generic across fleets and tools, support multi-platform switching, and avoid depending on a single data source.

5. Case Review (How Others Did It)

• Rook tested and failed: On TaskMarket (USDC escrowed on Base), Rook won all 192 tasks, but settlement was blocked by prerequisites—missing signed wallet, unmet withdrawal threshold, manual-review delays—leaving a zero available balance by night’s end. Lesson: configure wallets and withdrawal thresholds before accepting jobs.
• Toku’s low-unit-price pain: On a $15 order, Toku pocketed only $4.50 after platform take rates and payment-channel friction ate the rest. Most demand appeared to come from 0ai-Supervisor, not real paying humans. Lesson: monitor agent wallet spend to prevent search costs from exceeding net profit on a single task.
• dealwork’s auto-settlement model: Supports human-AI hybrid bidding and settles via Stripe; all seven contracts were auto-approved and paid within roughly a day. This shows some platforms are solving automated验收, leaving room for a manual QA layer to step in as a supplement.
• Upwork’s “human-AI pairing” test: Upwork found pairing humans with AI agents boosted task completion rates by 70%, but 14% of work was underpaid due to grader misjudgment. That points to “manual acceptance QA” as an investable service—charging a spread by helping clients collect more through higher验收 accuracy.
• Cost-cap tool practice (inferred): Tools like CashClaw poll task boards and kill processes the moment an agent’s spend exceeds a preset limit (say, $5 per task), avoiding mindless burning. This is becoming standard risk-control for agent fleets.

6. Dual-Track Executability

Cross-border: Viable. Major settlement rails (Stripe, Coinbase x402, Circle) already support stablecoin or USD auto-payments, letting you serve the global developer community directly and acquire users via GitHub or Twitter.Domestic (China): Not viable. China lacks a mature public AI-agent task marketplace, and cross-border payments plus on-chain settlement face compliance and liquidity barriers. This track isn’t ready for launch.

Original · Trends.vc: Read original →