Seven Criteria for Selecting an AI Agent for Real Work
Seven Questions from the Agent Passport
These seven questions come directly from the eight fields of the agent passport defined for every launch: task, input, process, output, limits, permissions, environment and verification (more on the pricing page). An agent with a completed passport answers every question; an agent without a passport is a polished presentation with unpredictable behavior.
Questions 1–3: task, data, stopping
1. What is the task and the result criterion? The guiding principle is to define the task, input data, expected artifact, acceptance criteria, limits, and decision-maker before launch. These inputs make it possible to assess the price and scope.
2. What data does the agent receive? The trust model requires least privilege: every source and tool is connected separately, access to other data must not be implied by the interface.
3. Where does the agent stop and ask a human? Human control is mandatory: disputed decisions stop and wait for confirmation, sensitive actions are explained before consent. Defined stopping points make the process manageable.
Questions 4–5: sources and the boundaries of claims
4. How is the conclusion linked to the source? In a verifiable agent every conclusion is traceable: the work journal records inputs, decisions and sources; the result can be followed back to the source data. Examples: the RAG navigator answers with exact quotes and references, the database consultant — with a link to a source table, the requirements analyst ties every requirement to a source.
5. What is the agent's defined scope of claims? The limits must be named in the passport: a sample is not market statistics, minutes require human verification, a RAG answer does not constitute legal advice, an audit is not a certification. Recording the limits in the passport makes them visible to every participant.
Questions 6–7: environment and artifact
6. In which environment does the data run? Cloud, team or local — the environment is defined before launch together with the retention period and the list of processors. For sensitive data this is the decisive question: local processing architecturally excludes passing to third parties.
7. In what form is the result returned? A strong result is a verifiable artifact: a document, a draft, a table, a report with quotes. The artifact can be opened, edited and passed on. If the result exists only in the chat with the agent and nowhere else — it can neither be checked nor used.
How to use the checklist
Use these questions to evaluate both ready-made products and a custom implementation project. The next steps are a demo launch to test the claims, the method to define the criteria, the security principles for the environment and access. If questions remain after the demo, return to the assessment before making a purchase decision.
Adapted from the article: agentseffect.com.