Sparround

What not to automate

The most expensive mistake with an agent is not technical — it is automating the wrong work. There are categories where, however well the agent performs, automating is the wrong decision:

  • Irreversible operations — production deploys, deleting data, payments, signing contracts
  • Decisions with legal or regulatory consequences — credit decisions, hiring, disciplinary action
  • Outbound communication — a message to a customer, a public post, official correspondence
  • Calculations that must be exactly right — accounting, regulatory reporting
  • Operations that widen access — granting permissions, creating keys, adding users

This does not mean "the agent must never touch these". The right pattern is the prepare / execute split: the agent prepares the draft, a human approves and executes.

  • instead of deploying → have it write the release notes and a risk list
  • instead of messaging a customer → have it draft the reply
  • instead of deleting from a database → have it produce the list of rows that would be deleted
  • instead of granting access → have it collect who asked for what and why it is needed

The split is also practical: preparation eats 80% of the time, approval the other 20%.

QuestionIf the answer is yesConclusion
Can a wrong result be undone?YesSafe to automate
Does a wrong result leave the building (customer, public, regulator)?YesRequires human approval
Is checking the result cheaper than doing the work?YesLet the agent prepare and a human check
Is the result checkable at all?NoDo not automate

The threat model in the docs points the same way: the primary focus is the "honest-but-wrong agent" — not malice, but a wrong command issued with the right intention. That is the correct mental model for automation decisions: the agent does not mean you harm, it just sometimes gets it wrong — quickly, and at scale.

Practice. Add two columns to the role matrix from the previous topic: "reversible?" and "does the result leave the building?". Then re-assess every row. Done means: you have downgraded at least one row from "fully automatic" to "the agent prepares, a human approves".

📚 Sources and documentation