AI coding agents that need approval before production
Let coding agents work freely in branches and staging, and require a person to approve anything that reaches production. With NobleCloak, policy sets which tools an agent may call alone, which wait for an approver, and which are denied outright, such as force-pushing or reading secrets. Every decision is recorded.
100 scenarios in 5 areas
These are scenarios for how teams govern AI agents in Software Development & Engineering. Some depend on capabilities still in development. Products named in a scenario are systems your team runs, named to make the example concrete. The policy applies to the AI agent's calls into them.
Source Control & Code Review20 examples
Repo deletion confirmed
Archiving or deleting a repository through an AI session stops for an in-session confirmation that names the repo and its open PRs.
Making a repo public held
Changing a repository's visibility to public asks for confirmation, since one click can publish proprietary code.
Security review for sensitive paths
PRs touching the auth module should route approval to a named security reviewer.
Coding agent cannot merge its own PR
A Cursor or Claude Code agent can open pull requests through the GitHub MCP tools, but the merge tool is denied for every AI session, so a human always merges.
Read the full exampleForce-push denied everywhere
The GitHub force-push and branch-delete tools are denied org-wide for AI, so an agent cleaning up history cannot wipe a teammate's commits.
Branch protection edits by role
Only the Platform Admin role can see the tool that changes branch protection rules; regular engineers' assistants never get it listed.
Contractors limited to their repos
Contract developers' assistants should reach only the repositories named in their statement of work.
Review bot comments, never approves
An AI code-review agent can post review comments but the "approve PR" tool is outside its allow-list, so required-review rules mean a human approval.
Every agent commit traceable
The audit ledger ties each AI-authored commit and PR to the engineer whose session produced it, which a fintech uses for change-management evidence.
Bank core repos read-only to AI
At a bank, the repositories that hold core-banking integration code are read-only for AI sessions; writes go only through humans.
Sub-agent refactor scope
A refactoring sub-agent spawned by the main coding agent can edit files in a working branch but does not inherit the parent's ability to open PRs or comment.
Org membership changes kept human
Inviting or removing members from the GitHub org is not entitled to any AI principal, so assistants never see it.
Partner agent opens issues only
A vendor's support agent may file issues in a shared integration repo via A2A, but cannot push code.
Staging repo vs. production repo
An engineer's agent has allow on the sandbox GitHub connection and ask on the main org connection.
PR spam rate limit
A per-user quota caps how many PRs an agent may open per hour, preventing a runaway loop from flooding reviewers.
Code-review agent cost by team
Engineering managers see the token cost of AI review per team and per repo to decide where it pays for itself.
Agent cannot pull random git remotes
The coding agent can only clone from the company's GitHub org through the governed proxy; arbitrary remotes and paste sites are blocked.
Webhook edits denied
Creating or editing repository webhooks is denied for AI, since a webhook can quietly send every push to an outside server.
Inspect a new hire's GitHub reach
Before onboarding, the eng lead uses "inspect as user" to confirm a new engineer's assistant can open PRs but cannot touch branch protection.
Deploy keys not visible
The deploy-key management tool is not entitled to the engineering org at all, so it never appears in any assistant's tool list.
CI/CD & Releases20 examples
Production deploys confirmed
When an engineer's assistant triggers a production deploy through the GitHub Actions or Argo CD tool, NobleCloak pauses for confirmation with the commit and target environment.
Rollback is allowed, roll-forward asks
Rolling back a deploy is allow so on-call can recover fast, while promoting a new version asks for confirmation.
Pipeline config changes confirmed
Writes to workflow files under `.github/workflows` via the agent should ask, since they control what runs with credentials.
Feature flag flips gated
Toggling a LaunchDarkly flag in production asks for confirmation; toggles in development run freely.
Named approver for prod at a health system
A healthcare SaaS routes AI-initiated deploys to the HIPAA environment to a named release engineer, who approves each one.
Staging deploys run free
The same deploy tool is set to allow on the staging connection, so agents iterate quickly without prompts where nothing customer-facing changes.
Release managers own prod
Only the Release Manager role can see the production promote tool; other engineers' assistants can deploy only to preview environments.
Change-window freeze
A bank blocks AI-triggered production deploys during the quarter-end change freeze.
CI secrets settings denied
Editing CI/CD secret variables in GitHub Actions or GitLab is denied to AI sessions, so an agent cannot add an exfiltration secret to a pipeline.
Release evidence for SOC 2
The audit ledger records who approved each AI-triggered production release and the inputs, giving the SOC 2 auditor the change-approval sample directly.
CI minutes budget
A build agent that buys extra runner minutes is held to the approved CI vendor and a monthly budget, with large purchases approved.
Release-notes agent is read-only
A release-notes agent reads merged PRs and Jira tickets and drafts notes, but cannot tag releases or publish them.
App store submissions by role
Submitting a mobile build to the App Store is entitled only to the Mobile Release role.
Deploy agent spawns test runner
A deploy orchestrator spawns a smoke-test sub-agent that can hit health endpoints but cannot itself deploy or roll back.
Build agents can't reach the internet
CI agents fetch packages only through the governed proxy and approved registries, not arbitrary download URLs.
Pipeline cost per service
Platform engineering tracks agent tokens and tool calls per service pipeline to find which build automations are expensive.
Vendor deploy agent called in
A managed-service partner's agent is allowlisted to trigger deploys of its own integration component and nothing else.
Deploys only inside the release workflow
The prod deploy tool should be callable only from the approved release workflow, not from ad hoc chats.
Inspect on-call before rotation
Before an engineer joins the release rotation, the admin previews their effective policy to confirm prod promote shows as ask.
Deploy rate limit per user
Each engineer's assistant can trigger a capped number of production deploys per day.
Production Infrastructure & Databases20 examples
Destructive SQL gated
Running `DROP`, `TRUNCATE`, or unscoped `DELETE` through the database tool should require confirmation even on staging.
DBAs get prod writes with confirmation
The DBA role can run production write queries, each paused for in-session confirmation showing the statement.
Read the full exampleTerraform apply confirmed
When an agent calls the Terraform apply tool, NobleCloak stops for confirmation; plan runs without prompts.
Instance termination held
Terminating EC2 instances or deleting RDS databases asks for confirmation, naming the resource and its tags.
Kubernetes scaling allowed, deletes ask
The Kubernetes tool allows scaling deployments but asks before deleting namespaces or persistent volumes.
Large-row updates held
Updates expected to touch more than 1,000 rows should ask for confirmation.
Prod Postgres is read-only for AI
Engineers' assistants have allow on read queries against the staging Postgres connection, but writes on the production connection are denied.
AWS IAM changes denied
Creating IAM users, access keys, or policies through the AWS MCP tool is denied for all AI sessions.
Cloud spend by agents capped
An infrastructure agent that provisions resources is limited to the approved cloud account, a per-call cap, and a monthly budget, with anything above the cap held.
Snowflake PII tables invisible
At an insurer, the Snowflake connection that holds policyholder PII is not entitled to general engineering, so assistants never see it.
Query results PII-scrubbed
Customer SSNs and account numbers should be masked in query results returned to an assistant.
Migration evidence
Every AI-run schema migration in production is recorded in the ledger with the engineer, agent, and SQL, which the compliance team uses for change logs.
Per-connection targeting by region
The EU production database connection is invisible to US-based engineers' assistants to support data-residency commitments.
Cost of AI database queries
Data platform sees how many tokens and warehouse queries each team's assistants generate against Snowflake.
Infra agent sub-agents narrowed
A capacity-planning sub-agent can read CloudWatch metrics but does not inherit the parent's ability to resize clusters.
Cloud console tools blocked to contractors
A contractor's per-user override denies every AWS write tool while still allowing read-only describe calls.
Backup deletion denied
Deleting snapshots or S3 backup buckets is denied org-wide, protecting recovery points from a confused or hijacked agent.
Agent reaches only cloud APIs
The infrastructure agent's outbound traffic goes through the governed proxy to AWS and the internal APIs only.
Credit union core DB untouchable
A credit union keeps its core processor database invisible to every AI principal; reporting runs off a replica connection with read-only access.
Inspect an SRE's prod reach
Before granting SRE on-call access, the admin uses "inspect as user" to confirm which production tools will allow, ask, or not appear.
Secrets, Dependencies & Supply Chain20 examples
Secret rotation confirmed
The Security Engineer role can rotate secrets via the agent, with an in-session confirmation before each rotation.
New dependency additions held
Adding a dependency not already in the lockfile should ask the engineer to confirm.
Vault policy edits confirmed
Editing Vault access policies through the agent asks for confirmation from the security engineer making the change.
Secret reads denied to agents
Reading values from AWS Secrets Manager or Vault through MCP tools is denied for AI; agents get references, not plaintext.
Package installs through the proxy
Coding agents install npm and PyPI packages only through the governed proxy to the internal mirror, blocking typosquat downloads from arbitrary hosts.
Paid package registries capped
An agent subscribing to paid API or registry plans is limited to approved vendors and a monthly budget.
Secrets stripped from prompts
API keys and tokens pasted into a session should be redacted before reaching tool outputs or logs.
Third-party MCP servers allowlisted
Only reviewed MCP servers are entitled to the org; a developer adding an unknown community server through their IDE finds its tools are not listed.
Dependency-update bot scope
A Dependabot-style agent can open version-bump PRs but cannot merge or change CI configuration.
SBOM evidence for regulators
A bank's vendor-risk team pulls ledger records showing which agent introduced which dependency and who approved it.
Signing keys not entitled
Code-signing key tools are not entitled to any engineering role, so no assistant can sign a release artifact.
External scanner agent calls in
A security vendor's scanning agent is allowlisted to call the vulnerability-report endpoint via A2A, while other external agents are refused.
License-check sub-agent
A license-compliance sub-agent can read manifests but does not inherit the parent's ability to edit them.
Token creation denied
Generating GitHub personal access tokens or cloud API keys through an AI tool is denied org-wide.
Security lead overrides for incident
During a key-exposure incident, the security lead gets a per-user override to revoke tokens directly from their assistant.
Egress denied to paste sites
Agents cannot post code or logs to pastebin, gist, or file-sharing sites, since the only route out is the governed proxy.
Scan cost tracked
AppSec sees what the AI dependency-triage agent costs per repository per month.
Secrets access time-boxed
A break-glass grant to read a production secret should expire after one hour.
Ledger shows every secret request
Denied attempts to read secrets are logged with the agent, user, and prompt context, which surfaces prompt-injection attempts.
Inspect a contractor before key access
The security team previews a contractor's effective policy to confirm no secrets tools appear.
Incident Response & On-Call20 examples
Paging tools asked before broadcast
Triggering a company-wide PagerDuty incident or status-page update asks the responder to confirm the message first.
Customer comms held
Sending outage emails to customers through the support platform asks for confirmation before anything goes out.
Approval routed to incident commander
Risky remediation steps proposed by a responder's agent should go to the incident commander for approval.
Database failover confirmed
Triggering an RDS failover asks the responder to confirm, since it causes a brief outage of its own.
On-call gets restart rights
The On-Call role is allowed to restart services and clear caches through the agent without prompts, so recovery isn't slowed by confirmations.
Status page publishing by role
Only the Incident Commander role can publish to the public status page; other responders' assistants can draft only.
Incident timeline from the ledger
After an outage, the postmortem pulls every AI action taken during the incident window from the audit ledger, including denials.
Elevated access during incidents only
Production write access should be granted to responders only while an incident is open.
Triage agent is read-only
A log-triage agent reads Datadog and CloudWatch and summarizes in Slack, but has no remediation tools in its allow-list.
Remediation sub-agent narrowed
A diagnostic sub-agent spawned during an incident can read metrics but cannot inherit the parent's restart rights.
Vendor support agent in the loop
The cloud provider's support agent may receive incident diagnostics through an allowlisted A2A call but cannot act on the company's systems.
Incident tooling cost
The SRE lead tracks how much AI triage costs per incident to justify the tooling budget.
Security incident lockdown
During a suspected compromise, the security team switches all production write tools to deny for every AI session in one policy change.
Forensics agent cannot exfiltrate
A forensics agent analyzing compromised hosts can't reach the internet except through the governed proxy, keeping evidence inside.
Bank incident reporting
A bank documents for regulators exactly which AI actions ran during a cyber incident and who authorized them.
Junior on-call gets ask, seniors allow
A per-user override sets restart tools to ask for engineers in their first on-call rotation while seniors stay on allow.
Emergency cloud spend capped
An agent scaling capacity during a traffic spike can spend up to a fixed emergency budget before purchases are held.
Hospital EHR integration off-limits
A health system's incident responders' assistants can see integration-engine logs but the tool that replays HL7 messages to the EHR is denied.
Log output scrubbed of PHI
Patient identifiers in logs should be redacted before a triage agent posts summaries into Slack.
Inspect responder access before rotation
Before the on-call handoff, the admin uses "inspect as user" to confirm the incoming responder has restart rights and no deletion rights.
| Action | Who | Decision |
|---|---|---|
| open_pull_request | coding agent | allowed |
| deploy_staging | coding agent | allowed |
| merge_pull_request | coding agent | denied |
| deploy_production | release manager's assistant | needs approval |
| force_push | any agent | denied |
| read_secret | any agent | denied |
Approval requests go to the release manager in Slack or the web inbox (Teams and email coming soon).
Illustrative: names no package, import or API call
AI agents in Software Development & Engineering: common questions
Can an AI coding agent deploy to production?
Only through the gate you set. A common setup allows staging deploys and marks the production deploy tool as needs approval, so the agent prepares the release and a person decides when it ships.
Can an agent merge its own pull request?
Not if the merge tool is denied for coding agents. The agent opens and updates the pull request; a person reviews and merges it.
Can DBAs or SREs get more than other engineers?
Yes. Policy can be scoped to a role or to one person, so a DBA's assistant can request a production write, with approval, while other engineers' agents stay read-only. Inspect as user shows what each person's agent would get.
Can access be limited to a time window or a single workflow?
Time-boxed and workflow-bound grants are coming soon, and the scenarios that need them carry the Coming soon label. Until then access is set per tool, by role or by person.
Set this policy up for your own agents.
Start with your own agents, your tools and the approvers you already have.
Other functions
Finance & Accounting
Payments, the close, treasury, budgets and payroll.
IT & Security Operations
Identity, devices, threat response, SaaS admin and cloud.
People & HR
Hiring, onboarding, compensation, employee relations and training.
Sales & Revenue
Outreach, the CRM, quotes, contracts and renewals.
Marketing & Communications
Content, campaigns, paid media, social and the website.
Customer Support & Success
Ticket replies, refunds, privacy requests and escalations.
Legal, Risk & Compliance
Contracts, regulatory work, audit evidence, vendor risk and records.
Operations, Procurement & Facilities
Purchasing, vendors, inventory, facilities and projects.
Data, Analytics & AI Operations
Warehouse access, pipelines, agent-to-agent work, models and AI cost.

