Business software for AI agents: how to tell whether it holds up

An agent that clicks through screens or scrapes HTML is guessing. Check for a command line or API that returns JSON and exposes its own schema: which commands exist, which fields are required, which values are allowed. Errors should be machine-readable too, with a code that tells the agent whether to fix the call, wait or ask you.

Check coverage as well. Whatever only works in the UI, the agent does by detour or not at all.

The agent needs a key of its own that you can revoke at any time, not your password. Better still if the key is exchanged for a short-lived token instead of travelling with every request.

Some actions do not belong in an agent's hands: creating new keys, deleting the account, changing the password. The software should block them on the server for agent access, not merely advise against them in the docs.

An issued invoice cannot be taken back, only cancelled. The software should mark such steps as destructive and run them only with explicit confirmation. That confirmation has to live in the tool itself; a sentence in the prompt did not protect the database at Replit.

Before the agent changes anything important, it should be able to show you exactly what it would send: method, address, body. You then approve what you have seen rather than an intention in words.

Agents retry. If the connection drops, the agent tries again, and without protection the time entry or payment ends up in the system twice. Check whether creating a record accepts an ID of your own that lets the server recognise a retry, and whether batch operations succeed completely or not at all. Where a step must not be repeated, the error should say so.

Afterwards you want to know what happened. That means every action is a readable command you can run yourself, requests can be logged on demand without exposing the key, and you can see which key was active when. One key per agent makes this unambiguous.

Mistakes happen. What matters is how they are fixed: edit or delete drafts, correct bookings, cancel what was issued, with a record rather than a silent overwrite. Anything already billed should be locked, so that a correction in one place does not change a finished invoice.

When the agent hits a limit, such as a feature outside your plan or too many requests, the software should say so clearly: with a reason and whether waiting helps or you need to decide. An agent that mistakes a limit for an error keeps retrying or looks for a workaround.

Give every agent its own key with as few permissions as needed, a read key where reading is enough. Allow anything irreversible only when you ask for it, and look at the dry run before you approve. Agent access to Time Momentum is part of Pro.