Gates and guardrails: never let an agent spend blind

The four guardrails

1. 1

   A stated ceiling, in the first message

   A number for the session. Not a vague instruction to be careful. "Do not spend more than 500 credits in this session, and stop and tell me when you reach 400."

2. 2

   Approval before every spending step

   Read only calls are free and can run freely. Anything that spends should be described, costed and confirmed. Relax this once you trust a particular workflow, never by default.

3. 3

   A row cap on every enrichment

   The most expensive agent mistake is not running the wrong tool, it is running the right tool on ten times the rows you meant. A hard cap makes that impossible rather than unlikely.

4. 4

   A separate workspace for experiments

   If you are exploring what an agent can do, do it somewhere a mistake does not touch a real project or a real campaign.

Where the guardrails go

• In your first message, every session. The constraint block from lesson 3.3.2 is exactly this.
• In the skill itself. The published SOPs have gates built in, which is a large part of their value.
• In your own reusable instructions, if your client supports project level configuration, so you never have to remember.
• In the reconciliation habit: balance before, running total during, balance after. It is how you find out whether the guardrails held.

You have finished the Architect track

Scheduled and refreshing lists, connected to the systems your team works in, and drivable from a conversation or from code. If you have come through all three tracks you can now build a list nobody else has, defend what it cost, and run the whole thing without opening the app.