# Agent Diligence Questions

General information, not legal advice. Adapt to your agents, contracts and jurisdiction. From Agent Authority, a legal reference by Obehi Ota.

## The agent

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| What acts can the agent take for us, and what is the highest authority level we can configure? | What the agent can commit you to | The provider can’t list the acts, or the customer can’t set a lower authority level |  |

## Identity and credentials

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| Does each agent have its own identity, separate from human users? | Attribution: shows which acts were the agent’s | It signs in with a customer employee’s login |  |
| How are the agent’s credentials issued, scoped and stored? | The gap between permission and approval | Static API keys with full account access |  |
| When the agent is connected to a system, who approves its access, and is that approval recorded? | Evidence of what the principal approved | No one approves access, or the approval isn’t recorded |  |
| How long do credentials last, how fast does revocation take effect everywhere, and can one agent be revoked without disabling the whole account? | Authority that outlives the job; clause 8 in Agent Use Clauses | Keys never expire; revocation takes days or requires shutting down the account |  |

## Limits and enforcement

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| Which limits does the service enforce itself: amount per action, totals over a period, counterparty, category, terms? | Enforced limits hold; instructed limits don’t | Limits exist only in the agent’s instructions |  |
| Which actions are blocked outright, and which can pause for approval? | Shows whether a reviewer can release a prohibited action | A reviewer can override a hard limit, or a rejection leaves no record |  |
| Can the agent change its own permissions or settings? | Self-expanding authority | Yes, through a settings tool |  |
| Can the agent change vendor bank details or payment instructions, or approve its own actions? | Separation of duties | Yes, through the same account it orders with |  |

## Human approval

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| Which actions can require approval, and can we set the thresholds? | Whether the right person approves the exact action before execution | Approval is all-or-nothing |  |
| How is the approver authenticated and recorded? | A weak approval is weak evidence | Any team member can approve by chat reply |  |
| If the approval service is down when a check runs, what happens? | Controls should fail closed | The agent proceeds |  |

## Delegation

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| Does the agent call other agents, tools or services? Which ones? | Subagent authority and subprocessors | The list is unknown or chosen dynamically |  |
| How is each tool connection, such as an MCP server, authorized, and whose credential does it use with the system behind it? | The receiving service checks which account and operations the credential permits | The provider cannot identify the account or permissions for each connection, or the tool server forwards an incoming token to another service |  |
| Do subagents get their own narrower credentials? | Access should narrow as it passes down | They reuse the parent agent’s token |  |

## Logs and evidence

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| What does each log entry record, how long is it kept, and can we export it? | Logs are the evidence in any dispute | Kept 30 days; no export |  |
| Are logs tamper-evident, and will you preserve them on notice of a dispute? | Admissibility and preservation duties; clause 7 in Agent Use Clauses | No preservation commitment |  |
| Do logs identify credentials only by an identifier, and never record live tokens, keys, passwords or full card numbers? | Anyone who can read the logs could use the secrets in them | Logs can contain tokens, keys or card numbers |  |

## Security

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| How does the service handle instructions in retrieved content or external data returned by a tool? | Manipulated agents take unauthorized actions | It relies on the system prompt alone |  |
| Which risks has testing covered (for example, those in the OWASP Top 10 for LLM Applications, a list of risks rather than a test method), how were the controls tested, by whom, when was the last test, and what were the results? | Evidence of reasonable care | No structured testing, or a list of risks without test results |  |

## Contracts and liability

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| Will you commit to enforce configured limits and tell us when they fail? | Clause 5 in Agent Use Clauses | Only a general disclaimer |  |
| Is there a separate liability cap for losses from failed limits? | Clause 6 in Agent Use Clauses | All losses fall under the general cap |  |

## Regulation

| Question | Why it matters | Red flag | Response |
|---|---|---|---|
| Which regulatory role do you take for each use case, and what documentation do you supply? | EU duties depend on role | No position on roles |  |
| Does the agent tell the people it deals with that it’s AI, and whom it acts for? | Disclosure duties, and what the counterparty knew | No disclosure, or the customer can’t turn it on |  |

## In an acquisition, also ask

- Which agents can bind the company, and under which delegation of authority?
- Which counterparties have agreed terms that address agents?
- Any incidents, disputes or regulator contacts involving agents?
- Which agents make or influence consequential decisions about individuals?
- Has any agent acted outside its authority, and how was each case resolved?
- If the company kept a benefit or paid for an unauthorized transaction, what material facts did it know at the time? Did it choose to affirm while it knew facts were missing?
- Are agent limits enforced by system controls, and are agent logs complete for the period under review?
- What obligations have agents already created that the company is still bound by: open orders, renewals, commitments?
