Human-in-the-Loop

​

Overview

HITL (Human in the Loop) ensures the AI Agent obtains human confirmation before executing sensitive operations. Zeus provides Auto-Run modes and tool approval workflows, allowing users to flexibly control the Agent’s level of automation.

---

​

Auto-Run Mode

Auto-Run mode controls the level of automation when the AI Agent executes tools.

​

Three Modes

​

Tool Approval UI

When a tool requires approval, the frontend displays an approval card containing the tool name, parameters, current status (pending / approved / rejected / running / completed), and approve/reject buttons.

​

Approval Decision Logic

Approval decisions are based on the current Auto-Run mode:

• Run Everything: All tools execute automatically, no approval needed
• Use Allowlist: Only allowlisted tools execute automatically; other tools require approval
• Ask Everytime: All tools require user approval

​

Data Persistence

User Auto-Run mode settings and tool allowlists are persisted via the database. Each user has independent configuration. The allowlist has a unique constraint on user and tool name to prevent duplicates.

---

​

Trigger Conditions

HITL approval is triggered under the following conditions:

1. The tool is not in the allowlist
2. Auto-Run mode is “Ask Everytime”
3. The tool is marked as requiring confirmation

---

​

Interaction Flow

---

​

Interruption & Recovery

When a tool requires approval, the Agent’s execution is suspended and the state is persisted via the Checkpointer. After user approval, the Agent resumes execution from the breakpoint. User approval decision types: Rejected tools have a system message appended, explicitly telling the Agent not to retry that tool. For more technical details on interrupt recovery, see the Agent Loop and Runtime documentation.