PARMAN is testing the integration of agent-to-human task handoffs, allowing AI agents to plan tasks that require human execution for real-world steps. The system exposes `submit_human_task` and `get_human_task` via MCP with scoped API keys, requiring explicit human approval before any action.
- Requests specify work details, location, deadline, and required evidence, but do not automatically authorize spending or dispatching personnel.
- A submitted or approved task is distinct from a completed physical job; the team verified a zero-budget simulated workflow but has not yet demonstrated a real physical job.
- The team is seeking feedback from 3–5 builders on evidence requirements, handling of incomplete work, and identifying software-incomplete real-world steps.
PARMAN aims to gather input from agent builders to refine how agents handle human-completed results and international requests, with feasibility and costs agreed upon per task.