KilatBot Articles

Articles

AI agent approvals: control and limits

· · Updated

KilatBot asks for approval before running a command classified as risky. You see the command and choose whether to allow it. But different actions use different checks. Low-risk commands run without asking, and the website publishing tool has an important limit you should understand before enabling access.

This guide explains the controls you have. They do not catch every risk, so keep checking results.

Separate instructions from enforced limits.

“Ask me first” is a useful instruction. Do not treat it as a lock on every action. An agent being told to ask is different from a tool that stops until you answer.

In KilatBot, risky commands wait for approval. The Publish websites tool, however, does not stop to ask on its own. The Coder template starts with this tool enabled and is instructed to ask first. If you do not want the agent publishing websites, turn the tool off under Tools. See how approvals work with memory, schedules and tools in the agent overview.

Start with access needed for one job. Example: to review a writing draft, do not enable website publishing because you might need it later. Clear instructions and limited access serve different purposes. Use both rather than expecting either one to cover every risk.

Read the request, not just the button.

When an approval request appears, read the action shown. Check what it will change and whether that matches your task. If the purpose is unclear, deny it and ask for a simpler explanation.

The choices are Allow, Allow for this chat, and Deny. For an initial trial, choose a one-time approval when you understand the action. Do not grant broader permission just to avoid interruptions from later requests.

Example review questions: “Which file will change? Why is this step needed? Can you prepare a draft without changing the source?”

Your approval records your decision to permit an action. It does not mean the result will be correct or reversible. Check the result before assigning the next step. If a request contains unfamiliar commands, a confident explanation alone is not a reason to approve it.

Know which action you are approving.

Several paths differ. Purchases require you to select Buy; the agent cannot buy by itself. Making a shared page public also needs your action.

Website publishing is different: its tool does not have its own approval step.

Commands classified as low-risk run without asking. For work involving important data, keep a source copy and review enabled tools. An approval feature is not a reason to hand over the only copy of material you cannot lose.

When should you take over the browser?

If the live browser is available in your workspace, you can watch it in dashboard chat. Select Take control to drive it yourself; the agent pauses while you are in control. Select Give back when finished. The agent checks the page again before continuing.

The agent can ask you to handle a sign-in, captcha, or payment. Enter sign-in and payment information in the browser, never in chat. For a task started on Telegram, a sign-in request sends you to dashboard chat because that is where you take control.

Before giving control back, check the open page and the next intended step. Taking over is not blanket approval for everything that follows. Use Stop browser when you need to close it. Read the approvals and browser guide for exact controls.

What happens to approvals while you are away?

A scheduled task has nobody there to answer an approval request, so risky commands in it are blocked. Separate preparation from steps that need your decision.

Example: schedule a summary for review, rather than a chain of actions that can only finish after your approval. Try the real path before making it routine. If it stops, identify the step needing attention instead of expanding access to force completion.

Read next: How to schedule AI agent tasks throughout the day · Recurring tasks: daily data analysis, content drafts

Evaluate control with a small trial.

Use sample material and actions that avoid important data. Try denying a request, then check that the outcome matches your decision. Record anything unclear before expanding the job.

Count the time spent reading requests and checking results too. A task needing constant decisions may suit working alongside you better than an unattended schedule.

Start with clear boundaries.

Prepare one job, its required tools, and actions you will not permit. Get started.

Built to work for you

Put your agent to work.

Start from a template and shape it to your business.

Get started