Development of a system for transferring tasks between an AI agent and a human (Handoff)
Handoff is a critical point in a hybrid workflow. A bad handoff loses context, duplicates work, or leaves the person without the necessary information to make a decision. A good handoff is seamless: the person receives everything they need to continue.
Types of handoffs
AI → Human (escalation): The agent has reached the authority or confidence limit. It communicates the task with full context: what was done, what was attempted, why it stopped, and the recommended next action.
Human → AI (delegation): A person assigns a task to an agent. The task format is important: a clear goal, constraints, success criteria, and context. We develop templates for different types of tasks.
AI → AI (internal transmission): One agent completes a part → passes it on to another. Formalized context package: results, metadata, accumulated state.
Context packet on escalation
Minimum standard format:
{
"task_id": "...",
"original_request": "...",
"steps_completed": [...],
"current_state": "...",
"reason_for_escalation": "...",
"recommended_action": "...",
"urgency": "high|medium|low",
"data_links": [...]
}
Notification System
The user receives a notification via a priority channel with full context and action buttons. SLA: critical → 15 minutes, normal → 4 hours. If the SLA is exceeded, the issue is escalated to the next level.
Feedback Loop
After each human decision, feedback is provided to the agent: whether the agent escalated correctly, whether the recommendation was correct. The data is used to train the escalation decision model.