The Production Readiness Gate
A pass-or-fail gate across nine control groups. Partial satisfaction is failure.
No agent reaches production without satisfying every item that applies to it. This is a gate, not a scorecard; partial satisfaction is failure. Where a control cannot be met, the exception process in Section 18 applies—routing around the gate does not.
The gate is in two parts. 17.1 is the Level 2 minimum bar, which every agent must clear to reach production. 17.2 is the Level 3 enforced state, gated to higher tiers where the consequences of a control gap scale with the authority granted. Within each, items are grouped by control domain. An item with nothing in its final column applies to every agent, including T0.
17.1 Level 2 — The Minimum Bar #
DISCDiscovery and Inventory
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| DISC-1 | Registered in the agent registry with a named human owner, before deployment | ||
| DISC-3 | Every MCP server the agent reaches is on the approved list | Agents using MCP |
AUTHIdentity and Authorization
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| AUTH-1 | Authorization scope documented, minimized, and tied to a business function | ||
| AUTH-2 | Delegation chain documented for every agent this agent may invoke | Multi-agent deployments | |
| AUTH-3 | Deprovisioning criteria and process defined | ||
| AUTH-4 | Identity planes separated, or the collapse documented with a remediation date | T2 and above |
CREDCredentials and Secrets
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| CRED-1 | Credential lifetime defined and enforced by expiry, not rotation policy alone | ||
| CRED-2 | Credentials reachable from the execution context enumerated and bounded to declared scope | ||
| CRED-3 | Credentials bound to a named workload identity; no shared static keys | ||
| CRED-5 | Revocation tested for every credential type, with measured time-to-effect |
ISOExecution Isolation and Network Egress
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| ISO-1 | Egress default-deny with an enumerated allowlist, each entry owned and justified | ||
| ISO-2 | Trust-tier separation from privileged agents confirmed in the running environment | ||
| ISO-3 | Runtime isolation verified in the running environment, not asserted in design | ||
| ISO-4 | Package installation via an internal mirror inside the production VM program | T2 and above |
TOOLTool and Action Authorization
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| TOOL-1 | A single enforcement point that every tool call traverses before any side effect | ||
| TOOL-2 | Tool allowlist enforced outside the model; unknown tools fail closed | ||
| TOOL-3 | Velocity ceiling configured with thresholds derived from observed behavior | ||
| TOOL-4 | Blast-radius ceiling declared for records, spend, and systems touched | ||
| TOOL-6 | Agent-to-agent calls authenticated; token passthrough confirmed absent | Multi-agent deployments | |
| TOOL-7 | Decision records emitted to a store outside the agent’s own write scope |
DATAData, Memory, and Context Integrity
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| DATA-1 | Provenance recorded for every context element; external content distinguishable | ||
| DATA-1 | Retrieval entitlement enforced at the data layer, not by post-retrieval filtering | ||
| DATA-4 | Session isolation tested for cross-principal and cross-tenant bleed | Multi-tenant deployments | |
| DATA-5 | Output filtering active on every egress path, including tool outputs | ||
| DATA-6 | Adversarial retrieval test executed against this agent’s corpus and passed |
SUPSupply Chain
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| SUP-1 | Model, tool, and MCP versions pinned with provenance recorded | ||
| SUP-2 | MCP servers authenticated per OAuth 2.1, tokens audience-bound, descriptions reviewed | Agents using MCP | |
| SUP-3 | Third-party assessment complete, or its absence accepted by a named signatory | Third-party agents |
OBSObservability and Detection
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| OBS-1 | Action telemetry flowing to central collection with alerting on defined conditions | ||
| OBS-2 | Expected behavioral profile documented for this agent class | T2 and above | |
| OBS-3 | Telemetry correlatable with identity, network, and cloud control-plane logs | ||
| OBS-4 | Detection tested against a simulated compromise of this specific agent | T2 and above |
RESPIncident Response and Forensics
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| RESP-1 | Analysis capability for malicious artifacts named and reachable without external approval | ||
| RESP-1 | Kill switch tested for this agent, with measured time-to-effect | ||
| RESP-4 | Evidence preservation configured for the ephemeral runtime | ||
| RESP-5 | Third-party notification criteria defined with a named decision owner | T2 and above |
RETDecommissioning and Retirement
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| RET-1 | Decommission criteria and retirement owner recorded at registration | ||
| RET-5 | Evidentiary retention configured to outlast the agent |
GOVHuman Oversight, Governance, and Assurance
| ✓ | ID | Gate requirement | Applies to |
|---|---|---|---|
| GOV-1 | Every human-in-the-loop approval point documented, with volume and review time captured | With human approval | |
| GOV-3 | If running with reduced guardrails: inventoried, owned, isolation at or above production | Reduced-guardrail environments | |
| GOV-4 | Autonomy tier assigned, risk acceptance signed at the tier-appropriate level, review date set |
17.2 Level 3 — The Enforced State #
CREDCredentials and Secrets
| ✓ | ID | Gate requirement | From tier |
|---|---|---|---|
| CRED-1 | No standing credentials present in the execution context | T3 and above |
ISOExecution Isolation and Network Egress
| ✓ | ID | Gate requirement | From tier |
|---|---|---|---|
| ISO-1 | Every egress path patched on the production vulnerability cadence | T3 and above |
TOOLTool and Action Authorization
| ✓ | ID | Gate requirement | From tier |
|---|---|---|---|
| TOOL-3, TOOL-4 | Velocity and blast-radius ceilings configured to halt rather than alert | T2 and above | |
| TOOL-7 | Decision records tamper-evident and independently retained | T3 and above |
OBSObservability and Detection
| ✓ | ID | Gate requirement | From tier |
|---|---|---|---|
| OBS-4 | Automated halt or quarantine path configured, not routed to a human queue | T3 and above |
GOVHuman Oversight, Governance, and Assurance
| ✓ | ID | Gate requirement | From tier |
|---|---|---|---|
| GOV-2 | Control claims independently verified rather than self-assessed | T4 |