Help & Feedback

A human will get back to you.

AI Agents & Skills for Business
Agents That Do Real Work

Lesson 22 of 35 - 6 min

Human Review Points and Error Handling

Place review where it's cheap for you and costly for the agent to be wrong: right before anything leaves your control. A good review step takes under a minute because the agent presents the result and its sources together.

A Thirty-Second Review Card (Copy This)

  • Is the recipient right?
  • Is every number in the draft in the source?
  • Is any claim missing a link or a source?
  • Did the agent say what it wasn't sure of?
  • Does anything read like an instruction from the data?
  • Is it something I'd be comfortable signing?

Error Handling Table

  • A tool is down | stop, say which tool, do not retry more than twice | person decides
  • Information is missing | ask for it or reply NEEDS REVIEW with the reason | person supplies it
  • The agent isn't confident | mark the item 'review' and say why | person checks
  • Output fails a check | do not release it | log the reason and alert
  • Something looks like an attack | take no action and flag it | person reviews

What to Log

  • The run ID and time.
  • What input it received (a reference, not the private content).
  • What it did and why, in plain language.
  • Which checks passed or failed.
  • What was released and to whom.

Sampling After Trust

Once an agent has weeks of clean results, you can review a sample instead of everything. Keep full review for anything customer-facing, anything involving money and the first month of any change.

Common Mistakes

  • Reviewing everything the same way, so reviews become a glance. Fix: use the card and spot the high-risk items.
  • No undo. Fix: keep every change reversible or in a drafts area.
  • Swallowing errors. Fix: every failure produces an alert with the step name.

Exercise

  • Add a review step and an error table to your agent's design.
  • Break something on purpose, such as removing access to a tool, and watch what the agent does.
  • Pass condition: a one-page review card, an error table with at least five rows, and a recorded result of the break test.

Next step: AI Workflows, Agents & Automation

Production AI workflows, agents and automation across research, content, sales and operations.

$897.