Development

Approval gates and human handoff

• A Confidence Threshold That Sends Weak Tickets to People
• Adding Human Handoff to an Agent: Design and Metrics
• Approval Gates for High-Risk MCP Tools
• Effect Classifiers for Tool Calls: Read, Write, Money
• Giving an Agent a Billing Tool: Approval Gates
• Giving an Agent a Read-Only Finance Tool First
• Routing a Request to a Person or a Model: A Decision Rule
• What the Approval Screen Shows: Payload, Reasoning and Expiry
• When a Routed Request Comes Back
• When an Agent Stops and Escalates: A Decision Table
• Who May Stop the Agent: Approval Gates and Kill Switches
• A Backlog Cap for Human Review: Setting It from Actual Volume
• A Confidence Gate Before Escalation: Thresholds from Golden Set
• A Handoff Note the Next Agent Can Act On: Fields and Limits
• A Pending-Call Table: Schema for Escalated Writes
• A Pending-Call Table: Schema for Queued Approvals
• A Queueing Model for Human Review: Arrival, Service, Wait
• A Refund Desk Runbook: Who Approves and What to Log
• A Reopen Counter for Handoffs That Come Back
• A Resume Token the Agent Can Carry: Encoding the Escalation State
• A Service Level for Human Queues: Time to First Response
• A Weekly Review of Returned Handoffs: Who Reads It and Why
• An Approval Gate Before the Write: Queue, Timeout, Audit Row
• Contract Tests That Catch a Leaky Approval Gate
• Giving an Agent a Write Tool: Approvals and Audit Trails
• Handoff Summaries the Person Actually Reads: Fields and Length
• Hashing Arguments at Request Time So a Resume Cannot Change Them
• Items That Resist Automation: A Routing Table for Human Review
• Kill Switches for Agent Tool Calls That Fire Mid-Run
• Measuring How Often an Agent Asks a Human: A Weekly Review
• Reading an Escalation Spike: When Humans Are the Bottleneck
• Resuming an Approved Call: Preconditions, Idempotency, Failure
• Reviewing Agent Pull Requests: Checklists, Tests, Gates
• Routing by Intent, Not Score: A Table Mapping Tickets
• Shadow Mode for a Finance Write Tool: Run Writes, Keep Results
• Shadow Mode for a Money Tool
• Sizing the Human Queue a Confidence Gate Creates
• The Approval Queue Table: Schema and Expiry
• The Escalation Payload: Fields a Human Needs to Act
• The Handoff Record: Fields to Keep When a Person Takes Over
• The Reviewer's Two Minutes: Fields for a Fast, Safe Approval
• The Run State Table: What an Agent Loop Persists
• The Tool-Call Diff: Showing What Changed Since the Last Approval
• When a Gate Should Defer to a Person: Rules for Escalation
• When a Human Edit Teaches the Router: Updating a Rule
• When a Routed Request Comes Back: Reopening Closed Handoffs
• When Nobody Answers: Cancel, Retry or Escalate on Expiry
• When the Agent Proposes a Backfill: Batching and Throttling
• When the Agent Proposes a Refund: Approval Queues
• When the Floor Fires on Good Tickets: Reviewing Overrides
• When the Human Edits the Payload: Reconciling Changed Approvals
• Who May Approve: Roles, Delegation and a Signed Identity