PRACTICAL CASE · 07

How to Give an Agent Autonomy Without Losing Control

Level: intermediateTime: 8 minutesOutcome: a permissions and limits matrix

Automation does not break when the model gets a word wrong, but when it sends an email, publishes a post, or spends money without verification.

Situation

Divide actions into reading, preparation, and execution. Reading can be allowed automatically. Preparation creates a draft. Execution requires the owner’s approval and displays the recipient, amount, text, and validity period.

What to do

Add a daily budget, timeout, maximum number of retries, and an emergency stop. Log not only the error but also the reason for the decision. Every tool should have an allowlist of parameters.

Verifying the result

Test: ask the agent to prepare an email, send the email, and repeat the send. The first request creates a draft, the second triggers approval, and the third is blocked as a duplicate.

What usually breaks

The statement “external action was not performed” must be true. If the action actually occurred but the report conceals it, the system is unfit for use.

← All guides · Lab glossary →

We publish what works for us—and implement the same solutions for your business. We design AI automations, Telegram bots, chats, and AI agents for real-world processes. Discuss your project →