Guardrails Are Not Optional
Constraints are what separate working systems from expensive experiments
People get excited about AI agents and immediately try to give them maximum freedom. “Be creative. Do whatever you think is best. Surprise me.” And then they are surprised when the agent does something stupid.
Agents Need Boundaries
An agent without constraints is a liability. It will make assumptions you did not authorize. It will take actions you did not expect. It will optimize for something you did not ask for.
Constraints are not limitations. They are what make agents reliable enough to trust with real work.
The Constraints I Use
Every agent I build has explicit rules about:
- What it can do. A specific list of allowed actions. Not “do whatever you need to.”
- What it cannot do. Explicit prohibitions. “Do not delete files. Do not modify configurations without approval. Do not send emails without review.”
- When to stop. Clear triggers for pausing and asking for human input. “If you encounter an error you cannot diagnose in 3 attempts, stop and report.”
- Run modes. My agents operate in defined modes: discovery only, read-only investigation, approved modification, or emergency cleanup. They cannot jump between modes without authorization.
- Timeouts. Every agent session has an automatic expiry. No agent runs indefinitely without check-in.
- Verification. After every modification, the agent verifies its work. Check the site still loads. Check the tests still pass. Check nothing broke.
The constraints section of a prompt should be at least as long as the instructions section. If you are telling it what to do but not what to avoid, you are building a system that will eventually surprise you.
I will help you build properly constrained agents for your specific workflows.