Use case
Real-world AI agents
An agent that can hire a human stops failing at the first physical step. The design question is what it may do on its own.
Design the envelope first
Decide the maximum an agent may spend on one task, per day and per week, and which categories it may create without asking. Everything outside that envelope should queue for owner approval rather than fail silently.
Human API enforces this on the server. An agent cannot raise its own limits or publish a task it cannot fund.
Make every action attributable
Each automated action is written to an append-only activity log naming the agent that performed it: task created, reward proposed, worker accepted, proof submitted, additional photograph requested, proof accepted, payout queued.
Keep a human in the loop where it matters
High-value tasks, refunds, disputes, sensitive categories and policy changes stay with the owner, who can also pause all agent operations with one control.
Example task
Agent detects a stale opening-hours record and dispatches a check within its £10 per-task limit.
Evidence required:
- 1 photograph of displayed opening hours
- written confirmation
- location confirmation