Internal Interview Script
The Internal Interview Script guides a structured conversation with the agent’s business owner (for a procurement agent, usually the procurement manager) and your own product, engineering and security team before an agent goes live, or when you review one already running. Each question comes with a sign of a good answer, a follow-up, and the legal issue it tests. Allow about an hour.
Start with the business owner: engineers know what the system can do, but often not what counterparties were told. Record each answer here, with who gave it, since the person who knows a fact is part of the evidence for it. Bring the agent’s mandate, and update it with what you learn, especially which control enforces each limit.
Business owner and the outside world
| Ask | Listen for | Follow-up | Legal issue |
|---|
| Who deployed this agent, and on whose behalf does it act? | A named person and one legal entity | Did that person have authority to approve it? | Who the principal is, and whether the delegation was valid |
| What have counterparties been told about the agent, and what do they see when it contacts them? | Specific documents, messages and labels | Does anything describe it as authorized to buy or sign without stating limits? | Apparent authority |
| Which of its actions can’t be undone: payments, signatures, deletions, public statements? | A short list, each with an extra check | What stops it before one of those? | Where limits and approval steps matter most |
Scope
| Ask | Listen for | Follow-up | Legal issue |
|---|
| Walk me through one real task this agent does, start to finish. | Specific steps, systems and people named | Where does a person see what it did? | What the agent actually does, as opposed to what the mandate says |
| Which outside systems can it change: order, pay, send, sign or delete? | A short, definite list | Is there anything it can do that it hasn’t done yet? | The range of acts that could bind the company |
Identity and credentials
| Ask | Listen for | Follow-up | Legal issue |
|---|
| What identity does the agent use in each system? | Its own client or service identity | Does it ever use an employee’s login or a shared key? | Attribution and the gap between permission and approval |
| What exactly can each credential do? | Named scopes, caps and expiry dates, with the service that checks each one | What request would that service refuse? | Technical permission compared with the mandate |
| Who issued each credential, and who can change it? | A named owner and a change process | Could the agent change its own permissions? | Who set the agent’s effective authority |
Limits
| Ask | Listen for | Follow-up | Legal issue |
|---|
| For each limit in the mandate, what stops the action if the model ignores it? | An order check before submission, or a payment check by the issuer for a card cap | Show me where that check lives. | Whether the limit is a control or advice |
| Which requests stop outright, and which go to a reviewer? | A hard limit blocks the action. A reviewable request waits for approval of that exact action. Both decisions are logged | What happens if the reviewer declines or the order changes? | Evidence, and the chance to reject promptly |
| How are monthly and company-wide totals enforced, including orders on invoice and orders placed at the same time? | One record of all commitments, checked before each order is sent, with pending orders counted | Could three vendors each accept orders within their own cap and together pass the monthly total? Is the card cap the only total? | Whether aggregate limits control commitments or only payments |
| Who receives counterparty confirmations and notices, and how quickly is each one read? | A monitored inbox with a named owner and backup; same-day triage | What happens to a confirmation that arrives Friday evening, or while the owner is away? | Missed rejection or dispute deadlines; what the company knew and did after notice |
| Are any limits enforced only in the instructions? | None, or a known short list | Which tool or service can enforce them before the action? | Exposure the company can’t see in its policies |
Content and instructions
| Ask | Listen for | Follow-up | Legal issue |
|---|
| What outside content reaches the model, including through tool results? | Named sources and the tools that return their content | Can a vendor or customer put text in front of it? | Prompt-injection exposure |
| How is outside content kept from acting as instructions? | Content is labeled as data and tool calls are checked after the model | Has anyone tested it with planted instructions? | Clause 14 in Agent Use Clauses, and reasonable care |
Approval
| Ask | Listen for | Follow-up | Legal issue |
|---|
| Which actions pause for a person, and what does the person see? | Clear thresholds; the full terms are shown | Can the agent go ahead if no one answers? | Whether approval is meaningful (clause 16 in Agent Use Clauses) |
| How is the approver authenticated and recorded? | Strong sign-in; name and time logged | Can anyone approve, or only named people? | Whether the approver had authority |
| For each approval step, what exactly does approving authorize: a message, accepting terms, a signature or a payment? | A written answer for each step | Does the approver know that? | What the approval proves |
Delegation
| Ask | Listen for | Follow-up | Legal issue |
|---|
| Does the agent call other agents or services? | A known list | Do they get their own narrower credentials? Can we trace who requested a quote and which system actually placed the order? | Attribution and authority for each handoff |
| Can a subagent reach anything the parent can’t through the parent’s delegation? | No. A separately authorized specialist service may hold its own credentials, but only under its own task-specific grant, checked against the company’s limits | How would we know if it did? | Authority growing as it passes down, or the combined system getting around the company’s limits |
Ending authority
| Ask | Listen for | Follow-up | Legal issue |
|---|
| How do we stop the agent right now? | A tested shut-off that takes minutes | When was it last tested, and how long until every token, key and cached session stops working? | Containment in an incident |
| When someone leaves or a mandate changes, what happens to the agent’s credentials and memory? | Both are updated in a defined process | Could the agent act on an old limit it remembers? | Stale authority |
| What happens to orders, offers and payments already in progress when we shut it off? | A list of unsent requests, outstanding offers and accepted orders, each with an owner | Who sends notice, when does it take effect, and which outstanding offers lapse? | Transactions in flight and notice to counterparties |
| Does shutting off this agent also stop any agents it started? | Yes, automatically | How would we confirm it? | Subagent authority after revocation |
Logs
| Ask | Listen for | Follow-up | Legal issue |
|---|
| What does each log entry record, and where is it kept? | Mandate version, relevant input details, retrieved source and version, proposed step, check result, approval and outcome | Can we export it, and is it tamper-evident? | Evidence in a dispute |
| How long are logs kept, and who can delete them? | Long enough to cover contract disputes; deletion restricted | Are any logs held only by a vendor? | Preservation duties |
Changes
| Ask | Listen for | Follow-up | Legal issue |
|---|
| What changes to the agent need approval before they ship? | Model, tools, credentials and limits all go through review | Who signs off on a new tool? | Whether the mandate still matches the system |
Related: Inside an Agent System · Agent Mandate · Agent Diligence Questions (for outside providers) · Agent Use Clauses