AI Agent Checker: Score Your Agent on the 12 Steps

Start with one specific problem.

• Write the job in one sentence.
• Name the person who receives the output.
• Define what done looks like.
• List what the agent must never touch.

The model acts as your agent's brain.

• Pick five real sample tasks.
• Run every candidate model on all five.
• Score accuracy and cost per task.
• Choose one and write down why.

Five parts, in plain language.

• Write each of the five parts in three sentences or fewer.
• Add one good and one bad example output.
• Review the instructions after the first week and tighten them.

Connect the information it needs.

• List the documents a person would need for this job.
• Remove duplicates and outdated copies.
• Connect sources through search instead of pasting them in.
• Give each source an owner.

Tools allow your agent to take actions.

• List the actions the job needs.
• Give each tool one purpose and a plain description.
• Start with read-only tools.
• Log every tool call.

This turns a chatbot into a system.

• Map which apps hold the data the job needs.
• Prefer official APIs.
• Use webhooks for events instead of polling.
• Keep credentials in one secured store, never in prompts.

Give your agent useful context across tasks.

• Decide what gets saved and who can edit it.
• Save every correction as a short lesson.
• Review long-term memory monthly and retire what is stale.

For complex tasks, let the agent break a goal into smaller steps.

• Ask for a written plan before the agent acts.
• Make each step small enough to verify.
• Require a check after each step.
• Run independent steps in parallel.

Protect your systems and information.

• List the worst things the agent could do.
• Block each one in code, outside the model.
• Validate inputs and outputs.
• Run new tools in a sandbox first.
• Prove each guardrail actually blocks.

Match the level of control to the risk.

• Sort every action into low, medium and high risk.
• Require approval for anything that sends, spends, deletes or changes settings.
• Ask for approval as a clear choice, not an open question.

Once the workflow works reliably, add triggers.

• Run the workflow by hand until it passes several times in a row.
• Add one trigger at a time.
• Send alerts only when something changes.
• Escalate anything left unacknowledged.

Measure, then improve.

• Keep a set of real test tasks.
• Re-run them after every change.
• Record failures in a ledger and fix the cause.
• Require two clean passes in a row before calling it done.