For operators about to put an AI agent in production

Don't flip that agent liveuntil it passes the gate.

Most AI agents look perfect in the demo and fail in production — and the failure is rarely the model, it's a missing guardrail. Answer five plain questions about each agent and get one verdict: READY, FIX, or DO NOT DEPLOY.

Get the Gate — $79one-time · instant download · yours to keep
Five deliverables · runnable
Readiness gate workbook
.xlsx
Runnable scoring engine
python
Facilitator Playbook
.docx
Go-Live Runbook
.docx
Worked example + config
json
Works alongside
Access Auditor · Reliability Harness · Red Team Kit
01.The Problem

The agent worked in the demo. Then it did something it couldn't undo.

5 controls

Approval gate, logging, rollback, bounded scope, escalation. Most agents reach production before anyone has checked all five — the guardrails get built last, in a panic, after something goes wrong.

Quiet failure

Agents fail quietly: a destructive action with no human in the loop, a refund loophole, a wrong booking at scale. The score looks fine right up until the incident.

Then ship

The fix is order: put the guardrails in first, then go live. This gate enforces that before launch — not after the postmortem.

02.See It Work

Cycle the controls. Watch one partial approval gate block an 86/100 agent.

Live demo

Tap a control to cycle READY → PARTIAL → MISSING. Toggle the destructive flag.

Gate fired. This agent can take a destructive action and the approval gate is not fully in place — it is DO NOT DEPLOY no matter how high the rest scores.

Readiness

86/100

DO NOT DEPLOY

Fix first: C1 Human approval gate on destructive actions

Same answers, same verdict, every time — deterministic and offline (as of 2026-06-25). Grades the guardrails you describe; it does not run or change any agent, and it scores no person.

03.The Runnable Engine

One verdict, computed the same way every time — in a spreadsheet or a script.

The same deterministic logic runs in the workbook and the Python engine. Here it scores a six-agent fleet and rolls it up to the worst agent. The teaching case is the refund agent — READY on four controls and 86/100 overall, yet DO NOT DEPLOY because it can take a destructive action with only a partial approval gate:

==================================================================
  AI Agent Go-Live Readiness Gate
  RedHub AI LLC · deterministic, offline · as of 2026-06-25
==================================================================

  Support tier-1 reply bot
  readiness 100/100   ->  READY
    OK  C1 Human approval gate on destructive actions     READY
    OK  C2 Full tool-call logging                         READY
    OK  C3 Tested rollback / recovery path                READY
    OK  C4 Bounded scope (refuses out-of-scope tasks)     READY
    OK  C5 Human escalation path                          READY

  Refund & billing agent  [DESTRUCTIVE]
  readiness  86/100   ->  DO NOT DEPLOY
    ~~  C1 Human approval gate on destructive actions     PARTIAL
    OK  C2 Full tool-call logging                         READY
    OK  C3 Tested rollback / recovery path                READY
    OK  C4 Bounded scope (refuses out-of-scope tasks)     READY
    OK  C5 Human escalation path                          READY
    !! GATE: destructive capability with no full approval gate -> DO NOT DEPLOY
    fix first: C1 Human approval gate on destructive actions

  Outbound email sender  [DESTRUCTIVE]
  readiness  19/100   ->  DO NOT DEPLOY
    XX  C1 Human approval gate on destructive actions     MISSING
    ~~  C2 Full tool-call logging                         PARTIAL
    XX  C3 Tested rollback / recovery path                MISSING
    ~~  C4 Bounded scope (refuses out-of-scope tasks)     PARTIAL
    XX  C5 Human escalation path                          MISSING
    !! GATE: destructive capability with no full approval gate -> DO NOT DEPLOY
    fix first: C1 Human approval gate on destructive actions

  Internal research assistant
  readiness  69/100   ->  FIX
    ~~  C1 Human approval gate on destructive actions     PARTIAL
    OK  C2 Full tool-call logging                         READY
    ~~  C3 Tested rollback / recovery path                PARTIAL
    OK  C4 Bounded scope (refuses out-of-scope tasks)     READY
    ~~  C5 Human escalation path                          PARTIAL
    fix first: C1 Human approval gate on destructive actions

  Calendar scheduling agent  [DESTRUCTIVE]
  readiness  92/100   ->  FIX
    OK  C1 Human approval gate on destructive actions     READY
    OK  C2 Full tool-call logging                         READY
    OK  C3 Tested rollback / recovery path                READY
    ~~  C4 Bounded scope (refuses out-of-scope tasks)     PARTIAL
    OK  C5 Human escalation path                          READY
    fix first: C4 Bounded scope (refuses out-of-scope tasks)

  Knowledge-base summarizer
  readiness  89/100   ->  FIX
    OK  C1 Human approval gate on destructive actions     READY
    ~~  C2 Full tool-call logging                         PARTIAL
    OK  C3 Tested rollback / recovery path                READY
    OK  C4 Bounded scope (refuses out-of-scope tasks)     READY
    OK  C5 Human escalation path                          READY
    fix first: C2 Full tool-call logging

------------------------------------------------------------------
  FLEET: DO NOT DEPLOY   (ready 1  /  fix 3  /  do-not-deploy 2  of 6)
  mean readiness 76/100 (context only; posture = worst agent)
  start with: Outbound email sender
------------------------------------------------------------------
  Grades guardrails you describe; does not run or change any agent; scores no person.
  Not legal or security-audit advice.
04.The Standard

Three operating principles behind the gate.

Gate then ship

An agent that can take an action it can't undo, with no human in the loop, is one bad input away from a public incident. So the approval gate dominates: a destructive capability without a full human approval step is DO NOT DEPLOY no matter how high the rest scores.

Every control, no averaging

Approval gate, logging, rollback, bounded scope, escalation. An agent is READY only when every control is in place — a strong score can't paper over a missing guardrail, because the failure isn't a lower average, it's a hard stop.

Scores guardrails, not people

Every check is about the agent you describe. It never connects to or runs your agent and never scores a person — same answers, same verdict, every time, in the workbook and the engine alike.

05.What This Is — And Isn't

A go/no-go judgment about an agent you control — not a security clearance.

What it is

  • A five-control gate that decides if an AI agent is safe to flip live.
  • A structured way to make the go/no-go call deliberately, before launch, instead of by vibes.
  • Deterministic and offline — the same answers always give the same verdict.

What it isn't

  • It scores the guardrails you describe, not people, and never connects to or runs your agent.
  • A readiness aid for the guardrails you describe — not a security audit, a penetration test, or legal advice. It does not connect to, run, or change any agent, and it scores no person.
  • A READY verdict is a readiness signal, not a guarantee — pair it with your own security review and counsel where the stakes warrant.
06.Who It's For

Anyone about to put an AI agent into production.

  • Founders and operators wiring an agent into real workflows.
  • Ops and RevOps leads who own the go/no-go decision.
  • Agencies setting a safe-to-launch bar for client agents.
  • Teams who want the guardrails in place before launch, not after.
08.Common Questions

The questions operators actually ask before deploying an agent.

Score it on five operational controls: a human approval gate on destructive actions, full tool-call logging, a tested rollback path, bounded scope, and a human escalation path. The AI Agent Go-Live Readiness Gate rates each READY, PARTIAL, or MISSING and returns one verdict — READY, FIX, or DO NOT DEPLOY. An agent is READY only when every control is in place; one missing guardrail is enough to hold the launch.

Put the guardrails in
before you go live.

One purchase, lifetime access, 12 months of updates. $79, once.

A readiness aid for the guardrails you describe — not a security audit, a penetration test, or legal advice. It does not connect to, run, or change any agent, and it scores no person.

Sold by RedHub AI LLC · Secured by Stripe · redhub.ai