Skip to main content
Free, no sign-up

Should you build that agent?

Most proposed agents should start as governed retrieval knowledge bases. This checklist tells you which ideas look agentic, which should become a safer knowledge layer first, and what control burden they may create.

The Test

List what you are planning to build

One or two words each. Then answer the checklist for every one. Two questions decide whether it is an agent. The rest size the control burden before it is safe to build. Nothing is sent anywhere, nothing is stored, and there is no sign-up.

  1. Is it an agent?

    Either one of these justifies it. Neither, and it is retrieval.

    1.Does what it does next depend on what it just found?

    The path through the work is decided at run time. If you can draw the whole flow in advance, you can write the whole flow in advance, and then it is a script.

    2.Does it need to run when nobody is there, and work out what to do when it does?

    Both halves are required. A schedule alone does not justify an agent, because a scheduled script runs unattended too. What makes it agentic is that the choice of what to do is not fixed in advance.

    Governance Preview

    None of these justify an agent. They decide what has to exist before whatever you build is safe.

    3.Does it reach two or more systems, under one identity?

    This cannot justify an agent: a scripted integration does exactly this. What it decides is how wide the allow-list has to be, and whose identity the thing acts under.

    4.Does it enforce a rule, and does the record of that have to stand up later?

    This cannot justify an agent either: a policy engine refuses and logs the refusal without a model involved. What it decides is how defensible the record has to be.

    5.Does it change anything in a live system?

    The one that stops a build. Anything that writes needs a control model agreed before it is built, not retrofitted after someone asks how it was allowed to.

    6.Does it access customer, employee, financial, health, or other sensitive data?

    Sensitive data changes the review burden even when the system only reads. Access, retention, redaction and evidence all need to be agreed.

    7.Would it act without a human approving each action?

    Autonomy is not only whether something runs on a schedule. The question is whether a human approves the action before it lands.

    8.Would a mistake create operational, financial, legal, reputational, or safety impact?

    The same technical action can be low-risk in one setting and high-risk in another. Impact decides how much evidence and fallback planning is needed.

    9.Would it act under a shared, admin, or human user identity?

    Identity determines what the system can touch and who is accountable. Shared or privileged access is a control problem before it is an agent problem.

Result

Answer the checklist for every agent to see the result.

The case people argue about

Onboarding that answers questions is retrieval, not an agent, however many systems it reads from. Onboarding that tracks where someone has got to, notices what they have not covered and adapts is stateful, which is agentic, and it needs memory, which carries its own governance problem. Same word, two different products.

The test is the public part of the audit

Knowing which agents are justified is the easy part, and it is the part we give away. The Knowledge Audit turns the no-results into a governed retrieval knowledge base plan, then adds the Governance Navigator review, tool inventory, control model and eval set where an agent is actually justified.

See the agent work
Talk to Us