AI Agent Capability Inventories: Knowing What the Delegate Can Really Do
How to keep a practical inventory of an AI agent's tools, permissions, model lane, context sources, review needs, and known limits before routing real work to it.

An AI agent is often introduced by what it appears to do: answer questions, draft messages, update files, browse pages, open tickets, or prepare code changes. That description is useful for a demo, but it is too loose for operations. A delegate that can browse a knowledge base is not the same as a delegate that can browse the open web. A delegate that can prepare a record update is not the same as one that can apply it. A coding agent that can edit a sandbox branch is not the same as one that can merge into a protected branch.
A capability inventory records what a particular agent lane is intended and authorized to do, what it has been tested doing, what context it may use, and what human review it needs. It sits before AI Agent Routing because routing only works when the routes point to known delegates. It also sits beside AI Agent Permissions and AI Agent Tool Contracts because capabilities are built from both authority and handles.
Capability Is Not One Thing
Teams sometimes describe an agent as capable or not capable at a whole category of work. That is usually too broad. The useful question is not whether the agent can handle support, research, finance, software, or operations. The useful question is which slice of that work has stable inputs, approved sources, bounded tools, clear success criteria, and a review path.
A research delegate may be capable of collecting source evidence from an approved corpus, summarizing tradeoffs, and marking uncertainty. It may not be capable of interpreting unapproved sources as policy or publishing final recommendations. A support delegate may be capable of drafting a reply from a current policy source. It may not be capable of promising refunds, changing account status, or inventing an exception. A coding delegate may be capable of reproducing a bug, editing a narrow file set, and running focused tests. It may not be capable of changing dependencies or deployment scripts without a separate approval.
This difference matters because success on one task does not establish competence across a whole category. The inventory gives the workflow a more precise language: this delegate can inspect, draft, compare, prepare, validate, escalate, or act under these conditions. Anything outside that shape needs a different route, a different permission, or a human owner.
Separate three questions: what the workflow is allowed to do, what its tools and credentials actually permit, and what evaluations show it can do reliably enough for this use. They may have different answers. An instruction to stay read-only does not itself remove a write permission, and a permission to write does not prove the agent can make a correct change. Check the effective access and enforce important boundaries in the tools or underlying systems, not only in the prompt.
For example, if the support drafting lane is authorized only to read tickets but its service account can also send replies, record that discrepancy and restrict the grant before using the lane for real tickets. Check the account’s effective permissions in the support system as well as the tool’s exposed functions; either layer can be broader than the intended task. OWASP’s guidance on excessive agency recommends minimizing functionality and permissions and enforcing authorization in downstream systems.
Make One Record Per Lane
Start with one workflow in one environment. An agent using the same model in a test repository and in production may have different permissions, evidence, and reviewers. Give each lane a stable name and record these fields:
- Purpose and owner: the task, intended users, accountable owner, and environment.
- Inputs and context: required inputs, approved sources, sensitive data limits, and known blind spots.
- Tools and authority: each tool or account, its effective read and write scope, and which actions require approval.
- Output and review: the artifact, evidence it must include, reviewer role, and acceptance criteria.
- Limits and escalation: recognizable stop conditions, the handoff destination, and any unavailable capabilities.
- Evidence and freshness: model and configuration version, evaluation results and their date, open gaps, last review, and next review or change trigger.
NIST’s AI Risk Management Framework calls for AI system inventories, clear responsibilities, and periodic review. A lane record is a practical way to capture details needed for routing; it is not a substitute for the organization’s broader system inventory or access controls.
For example, an inventory for a support reply drafting lane might say:
| Field | Example record |
|---|---|
| Task and scope | Draft replies to routine product questions in the test support queue; operations lead owns the lane. |
| Sources and tools | Read selected ticket fields and the current approved policy collection. No access to payment records. |
| Output and authority | Draft with links to the policy used and a note for missing facts. A support reviewer checks and sends the reply; the agent has no send or refund permission. |
| Stop and handoff | Escalate a missing or conflicting policy, a refund request, or an account-specific exception to the support reviewer. |
| Evidence | Record the configuration version, test set and results, evaluation date, and next review date. Until those are available, mark the lane as unverified rather than production-ready. |
This is an example, not a claim that any particular system has these controls or has passed those tests. Store links to evidence and configuration rather than credentials or unnecessary customer data. The record should be readable by the people who route and review work while access to sensitive details remains restricted.
For each capability claim, record the task and environment tested, configuration version, test date, number and kinds of cases, criteria for passing, results and failures, and a link to the evidence. For the drafting lane, separate ordinary policy questions from missing-policy, conflicting-policy, and refund-request cases. Count whether the draft used the right source and whether the agent escalated when required; review the actual failures as well as the totals. Choose acceptance thresholds before evaluation to fit the consequences and review process. A small or unrepresentative test set can reveal problems, but it cannot establish a dependable production failure rate.
Name The Handles The Agent Uses
A capability inventory should begin with handles. Which tools can the agent call? Which repositories, records, document shelves, queues, browsers, calendars, or messaging systems can it touch? Which of those handles are read-only, draft-only, or state-changing? Which ones are experimental? Which ones return structured evidence, and which ones return messy content that needs interpretation?
AI Agent Tool Contracts explains the tool side in detail. The inventory is where those contracts become operational. A tool named search is not enough. The inventory should say whether search reaches approved internal documents, public web pages, customer records, archived material, or all of those at once. A tool named update is not enough. The inventory should say whether it prepares an update for review, writes to a staging record, writes to production, or creates a request for another system.
The same habit applies to non-tool capabilities. If the agent can remember preferences, the inventory should say which preferences and where they come from. If it can use a browser, the inventory should say which pages are in scope and which actions are stop conditions. If it can create artifacts, the inventory should say what those artifacts must include. AI Agent Artifact Design becomes easier when the inventory already names the work product.
Record Context Sources And Blind Spots
Capability depends on context as much as tools. A delegate with the right tool and the wrong source can produce confident errors. A delegate with current sources and no permission to inspect the relevant record may stop correctly, but it is not capable of completing the task.
The inventory should name the sources the agent is expected to use and the sources it should avoid. A policy agent may use the current policy collection, not old training decks. A coding agent may use the repository, test output, and issue thread, not unrelated package examples from the web unless the task asks for them. A finance operations agent may use a ledger export and approval record, not a casual Slack summary. AI Agent Knowledge Bases covers the maintained shelf; the inventory says which shelf belongs to this lane.
Blind spots deserve the same attention. If the agent cannot see private attachments, live production records, historical exceptions, or customer-specific constraints, that is part of its capability profile. A blind spot does not always make the agent useless. It may simply mean the agent can prepare a draft and mark what needs human confirmation. But if the blind spot is hidden, the workflow will confuse partial work with complete work.
Include The Review Burden
An inventory that only lists what the agent can produce is incomplete. It should also describe what review the output requires. Some capabilities are valuable only when the review burden is reasonable. If the delegate saves ten minutes of drafting but forces a reviewer to spend twenty minutes reconstructing the evidence, the capability is weaker than it looks.
Human Review for AI Agents is the natural companion here. The inventory can describe the expected review surface: source trail, diff, proposed action, validation result, confidence boundary, or escalation note. It can also name the reviewer role. A policy reviewer, engineering reviewer, operations owner, and requester are not interchangeable. If every capability routes to a generic human queue, the inventory has hidden an operating cost.
Review burden also changes with maturity. A new delegate may require close review for every run. After the workflow has evaluations, observability, accepted artifacts, and a history of correct stops, some low-risk runs may need lighter review. That change should update the inventory rather than live as tribal memory.
Keep Limits As First-Class Data
The most useful capability inventories are honest about limits. They do not read like product marketing. They say where the delegate is brittle, which tasks are out of scope, which data is stale or partial, which tools are slow, and which conditions require escalation.
This is especially important for model behavior. AI Agent Model Selection helps choose a model lane, but the inventory should record what that choice means for the workflow. One model configuration may perform adequately on the lane’s classification tests but struggle with its synthesis cases. Another may meet the quality target but be too slow for a queue that needs quick triage. These are findings to test for the particular workload, not guarantees based on price or a model label. A browser-capable lane may be useful for gathering evidence but too risky for forms that can submit irreversible actions.
Limits should be written in operational language. Instead of saying the agent is “not reliable with ambiguity,” say that it must stop when the request lacks a record identifier, when two records match the same customer, when the governing source cannot be found, or when the proposed action would exceed a named permission boundary. Limits are useful when they can be recognized during a run.
Use The Inventory During Routing And Change
Capability inventory is not a document to admire once and forget. It should be used when a task is routed, when a permission changes, when a tool contract changes, when a new source is added, and when an incident or near miss reveals a gap.
AI Agent Change Management matters because capabilities drift. A model upgrade can change output quality. A tool can add a field or remove one. A knowledge base can move. A reviewer queue can become overloaded. A prompt can gain a new instruction that expands scope quietly. If the inventory is not updated, routing decisions will point to yesterday’s agent.
Before routing a task, compare its required inputs, sources, actions, and review needs with the record. If the record is stale, the required ability is untested, or the task crosses a stop condition, route it to an owner for review instead of assuming the lane is suitable. Link each capability claim to a dated evaluation or operational evidence, including failures and gaps, and use AI Agent Evaluations to test ordinary cases as well as missing inputs, conflicting sources, and boundary-crossing requests. A passing test set is evidence for the cases covered, not a guarantee for every future run.
For the example lane, that comparison could lead to three different decisions:
| Incoming work | Inventory finding | Routing decision |
|---|---|---|
| Routine product question with a current approved policy | Drafting is authorized and supported by current evidence; the reviewer is available. | Route for a draft and human review. |
| Customer asks for a refund | Refund requests are a named stop condition. | Route to the support reviewer; do not treat drafting permission as refund authority. |
| Routine question, but the policy collection or configuration changed after testing | The evidence may no longer cover the current lane. | Send to the owner for reassessment before relying on the affected capability. |
These decisions depend on the actual record and controls, not the labels alone. A lane can be suitable for a draft while the final action remains a separate human decision. If access is broader than authorized or an approval gate is missing, correct or disable the affected access rather than relying on a routing note.
Assign an owner to review the record on a schedule suited to the workflow and whenever a model, prompt, tool, credential, source, reviewer, or task scope changes. Compare the record with the actual configuration and relevant logs; resolve discrepancies or mark them open with an owner and due date. If a change invalidates prior evidence, limit routing to the previously verified scope or pause the affected lane until it is reassessed. AI Agent Access Reviews covers checking the effective grants behind the inventory.
The inventory also makes decommissioning easier later. AI Agent Decommissioning asks what the lane touched, what credentials it used, and what work needs redirecting. A current capability inventory answers much of that before shutdown becomes urgent.
A good inventory does not need to be long. It needs to be specific enough that a person can decide whether a task belongs in this lane. It names the handles, context, authority, review burden, limits, and owner. That small act turns an impressive demo into an accountable delegate. The team no longer asks, in a vague way, whether the agent can do the work. It asks whether this run fits the capabilities that have actually been designed.
Written By
JJ Ben-Joseph
Founder and CEO ยท TensorSpace
Founder and CEO of TensorSpace and the editor of Fondsites: an engineer working across software, AI, cybersecurity, and biosecurity who builds the site's games and tools.