Human-in-the-Loop

The inbox is the operator-facing governance surface for Stagent. It consolidates permission requests, agent messages, completions, and failures into a single review queue so agents never act outside the trust boundary you set.

Governance command center with pending, approved, and denied counts, notification queue with permissions and messages

Key Concepts

Mixed Feed — The default inbox view blends permissions, messages, completions, and failures into one triage surface with progressive disclosure.
Focused Tabs — Operators can narrow to unread items, permission requests, or conversational messages when they need a tighter queue.
Approval Controls — Permission cards expose Allow Once, Always Allow, and Deny actions directly on the notification. Always Allow patterns persist for future auto-approval.
Ambient Approval Toasts — Permission requests surface as edge-of-screen toasts on any page so you can approve or deny without navigating to the inbox.
Agent Handoff Approvals — Governance gates for async agent-to-agent handoffs ensure operators approve delegation chains before work transfers between agents.
Agent Questions — Agents can pause execution and ask for clarification. Your typed response is injected directly into the agent’s context so execution resumes with a clear human decision.
Budget Alerts — Notifications fire at 80% and 100% of configured spend caps, giving operators time to adjust before work is blocked.
Rich Content Rendering — Notification bodies render GitHub Flavored Markdown with code blocks, tables, lists, and links for detailed agent output.
Batch Context Proposals — When workflows generate multiple context suggestions, operators can review and approve them as a batch instead of handling each one individually.
Response History — Stagent preserves whether a request has been handled already, preventing duplicate decision-making.

Notification Types

Type	Action Required	Description
`permission_required`	Yes	Review a tool request and decide whether the agent can continue
`agent_message`	Yes	Respond to a clarification or provide missing direction
`task_completed`	No	Track successful completion and inspect the outcome
`task_failed`	No	Triage failures and decide whether to retry or intervene

How It Works

When a running task needs human input, Stagent creates an inbox item and waits for the response. That input might be a permission review, a direct answer to a question, or simply a signal that a run finished or failed.

The permissions tab is the highest-signal view when you want to supervise active execution. It strips away the general feed and shows only the decisions that currently gate an agent.

Governance command center Permissions tab with responded Write and SendEmail tool requests

Use Cases

Compliance-First Oversight

Product Team · Enforcing access policies across agent operations

Your compliance lead lives in the permissions tab. Read requests on safe paths get approved quickly, while outbound email or write operations require a closer review before the workflow can continue.

Interactive Research Session

Knowledge Worker · Collaborating with an agent on a financial analysis

The agent finds conflicting figures across two files and asks for clarification instead of guessing. You respond from the inbox, and the task resumes with a clear human decision attached to the record.

Learning to Trust AI

Non-Technical User · Gradually reducing oversight as confidence grows

You begin by reviewing every request manually. Over time, you promote safe actions to Always Allow, reducing noise without removing the audit trail or hiding what the system is doing.

Agent Integration — The canUseTool pattern that generates notifications
Monitoring — See agent activity in real time
Task Execution — Where task runs originate