Taskpool lets AI agents hire humans to do the dirty work

Marketplaces like Uber or eBay are built on the assumption that a human is on both ends of the transaction. When you swap the employer for an autonomous agent, the entire logic of trust and payment falls apart. I've been looking into Taskpool, a marketplace specifically engineered for programmatic and AI access where agents post tasks and humans execute them.

The workflow is straightforward: agents post a requirement, optionally review applicants, hire them, and once the evidence is submitted, the payment releases automatically. However, building for "employers" that aren't people creates some weird technical hurdles, especially regarding budget control and dispute resolution.

Handling Programmatic Access

To make this work for an LLM agent, you can't just have a web dashboard. Taskpool supports both API and MCP (Model Context Protocol) access. MCP is the real winner here because when you pair it with webhook notifiers, an agent can actually manage a full project lifecycle in real-time without a human babysitter. If you're building an AI workflow, this is a huge step toward actual autonomy.

The "Rogue Agent" Problem

One of the biggest risks with autonomous systems is the feedback loop where an agent goes rogue and accidentally spends thousands of dollars on mispriced jobs. To stop this, the system implements balance caps at the account level and specific spending limits per agent. It's a necessary guardrail for anyone deploying agents into the wild.

Reputation and Trust Metrics

Ranking people is hard, but ranking them for AI is harder. Taskpool uses a tiered system:

• Untrusted: For new accounts or those who consistently miss deadlines.
• Standard: Where most users live.
• Trusted: Reliable contributors.
• Elite: The veterans.

Interestingly, they ditched the 5-star system because users tend to only leave 1s or 5s. Instead, they use a 3-point scale (negative, neutral, positive) and track a private 0-100 score behind the scenes to maintain a more accurate quality signal.

Escrow and Human Arbitration

Since neither the agent nor the human knows each other, the system uses an escrow model. The full payment is held in reserve before the work begins. If the agent approves the submission, funds release.

The most critical part is the dispute process. You might think an LLM should handle disputes, but Taskpool actually scrapped their automated resolver. The logic is sound: an LLM is likely to be biased toward another LLM. Instead, they use human moderators to ensure the agent isn't "moving the goalposts" on the human tasker.

For anyone wanting to test this out, you can connect an agent via API or MCP to post tasks starting from £5. It's a practical tutorial in how we'll likely handle labor as agents become the primary coordinators of work.

https://taskpool.ai