Why prompt-only guardrails fail
A rule written into a prompt ("never spend more than $20") is a request, not a control. Models can misread it, be manipulated by input, or simply make a mistake. A real guardrail must sit outside the model and reject the action itself.
How Maqron enforces them
- Spending limits — checked by a database function before any cost is recorded; an action over the limit is blocked.
- Tool permissions — each tool must be granted to the business; ungranted tools cannot run.
- Approvals — in Co-founder mode every tool use waits for the owner. See human-in-the-loop.
- Evidence — every directive becomes an operating run with recorded actions and costs.
Prompt guardrails vs. enforced guardrails
| Question | Prompt-only | Maqron |
|---|---|---|
| Can the agent exceed a budget? | Yes, if it errs | No — blocked by the database |
| Can it use an unapproved tool? | Possibly | No — permission required |
| Is every action recorded? | Depends on logging | Yes — operating run |
| Who writes cost record entries? | Application code | Database functions only |
Spending limits for AI actions
Every tool cost the agent incurs is checked against your limits before it is recorded in the cost record. Read more on security and pricing.