An organization may have more autonomous agents running than its leadership realizes. An agent registry provides an inventory of active agents, their purposes, human owners, and operational exposure. It is a starting point for runtime governance, not a substitute for monitoring or action controls. section 3.8
Build and maintain the inventory around deployed agents:
- Record each agent’s owner, purpose, autonomy tier, approved tools, data scope, and whether it faces the public or operates internally. section 3.8
- Use a managed platform’s registry where available, but account for visibility limits and export a condensed view into your own inventory if needed. For self-hosted agents, a gateway such as IBM ContextForge can serve as a registry. section 3.8
- Do not rely on intake and procurement records alone. Use tools such as Zscaler and log monitoring to find agents that were never registered. section 3.8
- Connect identified agents to runtime monitoring and the control plane. Logs can record harmful actions after the fact; real-time checks are needed to prevent them. section 3.8
Table 3.18 Lifecycle artifacts and runtime controls for agentic systems
|
Lifecycle artifact
|
What Agentic AI requires
|
Level
|
|
GenAI Acceptable Use Policy
|
Policy as code[28]:
the same rules written so a system can check them before each action, instead
of a document people are asked to follow
|
1
|
|
AI Inventory
|
Agent registry[29]:
owner, purpose, autonomy tier, approved tools, and data scope. Filled by
scanning for agents already running, because teams do not always declare them
|
1
|
|
Risk Tier Framework
|
Tiered autonomy: what does the agent have access to and
not? Read only, internal facing, write capability, external facing, etc.
Controls must be stricter at each higher tier.
|
1–2
|
|
Model and system documentation
|
Agent manifest[30]:
the owner’s written statement of what the agent is for, what it may reach,
and what it must never do
|
3
|
|
Technical standards
|
Agentic identity, the gateway where every tool call is
checked (MCP gateway), policy engines and sandboxing (restricted environment
that agent runs on).
|
3
|
|
System configuration record
|
Memory governance[31]:
what the agent keeps between runs, how long it lives, who can delete it, and
how a deletion request reaches it
|
2, 5
|
|
Human review test results
|
Proof that the agentic run actually stops: the agent
pauses before an action it cannot undo, the reviewer can see what it already
did, and refusing ends the run
|
4
|
|
Request Record
|
Agent trace[32]:
each step, tool call, memory write and approval in order, and for every step,
which agent acted and who asked it to
|
5
|
|
Incident Response and Recovery
|
Kill switch and automatic limits: who can stop a running
agent, and limits on spend, repeated attempts, and loops that stop it without
waiting for a person
|
5
|
Compare registry fields with autonomy tiers and other lifecycle controls. section 3.8
