Change advisory boards have decades of practice reviewing software releases: code diffs, test results, rollback plans. AI agents break several assumptions built into that model. An agent's behavior is not fully determined at deployment time. Permissions can be invoked dynamically at runtime, tool access can be expanded through configuration rather than code, and the same agent version can act differently depending on the tools and data sources it is connected to at a given moment. NIST's AI Risk Management Framework treats this as a lifecycle problem, calling for documentation of risk decisions on an ongoing basis rather than as a one-time gate. A packet built for a traditional software release, focused on code changes and functional test results, does not capture the artifacts a CAB actually needs to evaluate an agent change: what the agent is now permitted to do, what tools it can call, and what governance checkpoints exist if it acts outside expected bounds.
What Belongs in the Packet: Identity and Permissions
Agent identity records, including service account bindings and credential scope, are the starting point. Unlike a human user account, an agent's identity often maps to a set of dynamically invocable permissions rather than a fixed role. The packet should include a permission diff that shows the prior authorized scope against the proposed scope, structured so a reviewer can see exactly what access is being added or removed. This directly addresses the "Excessive Agency" risk category named in OWASP's guidance for LLM applications, which describes harm resulting from agents granted more permission, functionality, or autonomy than their task requires. OWASP's recommended mitigation, least-privilege scoping combined with logging of tool invocations, translates directly into two packet artifacts: the permission diff itself and a corresponding tool-call log covering the review period.
What Belongs in the Packet: Tool Calls and MCP Connections
The Model Context Protocol has introduced a standardized, versionable interface layer for connecting agents to external tools and data sources. This creates a new class of auditable artifact that has no direct equivalent in conventional software change packets: the MCP server connection manifest. Because MCP-specific risks such as tool poisoning and unauthorized capability injection have been documented by security researchers, any change that adds, modifies, or removes an MCP server connection should be treated as a distinct risk review item rather than bundled into a general application change. The packet should include a version-controlled manifest of connected MCP servers, the capabilities each one exposes, and a record of what changed since the last approved state. This gives the CAB a concrete basis for assessing third-party tool risk instead of relying on a general description of "added integration."
What Belongs in the Packet: Policy Version History and Human Review
Runtime guardrails, such as allowed action lists, rate limits, and escalation rules, are policy artifacts distinct from application code. These should be version-controlled and diffable, with timestamps, so a CAB can see precisely what governance behavior changed between agent versions, not just what code changed. NIST's Generative AI Profile recommends that organizations define clear escalation and human review processes for systems capable of autonomous or semi-autonomous action. The packet should include a log of human review checkpoints tied to the specific change under review: who reviewed the proposed permission or policy change, what escalation path exists if the agent's behavior deviates from the approved policy after deployment, and how that deviation would be detected. Without this evidence, a CAB is approving a behavioral change with no record of who is accountable for monitoring its effects.
Practitioner note
A packet lacking a human review log is not incomplete on formatting grounds; it leaves the CAB with no accountable party to contact if the agent's runtime behavior diverges from what was approved.