30 checks before your agent goes live.
The AI Agent Pre-Production Checklist is a free 9-page PDF that reviews one AI agent at a time, from the identity it runs as to the alerts that watch it after launch. Each check tells you how to run it and what evidence to keep.
- Identity and permissions, 6 checks
- Tools and MCP servers, 6 checks
- Prompt injection and actions, 6 checks
- Data reach, 6 checks
- Monitoring after launch, 6 checks
- A go/no-go scorecard to sign
That did not go through. Check the email address and tick the box, then try again.
What the checklist covers.
An AI agent acts with an identity, calls tools, reads content it does not control and reaches data nobody mapped. Application reviews miss most of that. The checklist splits the review into five parts.
1. Identity and permissions
Whose rights the agent uses when it acts: its own identity, no admin rights, no shared builder credentials, scoped tokens, no secrets inherited by tool servers, a named owner.
2. Tools and MCP servers
Inventory every tool from the live server, pin each description and schema, stop the agent when a tool changes, verify annotations, pin package versions, expose only what the job needs.
3. Prompt injection and actions
Test every content source the agent reads, require a human for side-effect actions, allow-list outbound channels, treat tool results as data, cap steps and spend, keep secrets out of the model.
4. Data reach
List every source the agent can read or write, remove what is out of scope, enforce limits in the tool rather than the prompt, test cross-tenant access, define retention.
5. Monitoring after launch
Log every step with a session ID, notice when logging stops, alert on tool drift, on a disabled control and on unusual tool use, and reopen the review on every change.
How to use it on one agent.
- Pick one agent and one version. A new model, tool or permission is a new version.
- Mark every check Pass, Fail or Not tested. A check you could not run is never a Pass.
- File the evidence each check names. A config export, a log line, a test result.
- Apply the go/no-go rule. No go-live with an open Fail on identity or tools; every other Fail gets an owner and a fix date.
About the checklist.
What is in the AI agent pre-production checklist?
30 checks in five sections: identity and permissions, tools and MCP servers, prompt injection and actions, data reach, and monitoring after launch. Each check says how to run it and which evidence to keep, and a go/no-go scorecard closes the guide.
Who is the checklist for?
Security engineers, AI platform teams and CISOs who must decide whether an AI agent, including a home-built agent or an MCP server, is safe to put in production.
Does it cover MCP security?
Yes. Section 2 covers MCP servers: inventory from the live tools/list response, pinning tool descriptions and schemas, blocking a changed tool before the model call, verifying annotations such as readOnlyHint, pinning package versions and exposing only the tools the job needs.
Which frameworks does it map to?
The checks reference the OWASP Top 10 for Agentic Applications (2026 edition: agent goal hijack, tool misuse, identity and privilege abuse, agentic supply chain) and the security considerations of the Model Context Protocol specification.