AI agent guardrails written in a prompt can be ignored

Many AI agent guardrails are a paragraph of instructions. "Never share customer data." "Ask before issuing a refund." The agent reads them at the start of the task and follows them most of the time.

Most of the time is not always. Writing a rule in a prompt tells the agent what you want. It does not stop the agent from doing something else, and the longer the instructions grow, the easier a rule is to miss.

What the Technology Radar says

The Thoughtworks Technology Radar is a twice-yearly assessment of tools and techniques, based on what its teams see in their work. Its April 2026 edition makes the same point from two sides.

It places agent instruction bloat under Caution. The files that tell agents how to work grow as teams add conventions and rules, until the instructions become long and sometimes conflict with each other. Models pay less attention to guidance buried in a long context, so an important rule can be missed.

It places feedback sensors for coding agents under Trial. These are checks that give the same answer every time, such as compilers, linters and test suites, built into the agent's workflow. When a check fails, the agent goes back and corrects its work.

The first item describes the problem with rules kept in a prompt. The second describes the answer: a check that sits outside the agent, which the agent does not grade for itself. The Radar's examples come from software teams, and the principle holds for any agent.

What this means for a leadership team

A rule an agent can read is a request. A rule that is checked by someone or something independent of the agent is a control. Both have a place, but they are not the same, and a governance plan should say which is which.

If the agent grades its own work, nobody checked it.

How we work

Our own agents start every run from rules we hold ourselves, written down with the human whose judgment they carry. Each agent receives the rules for the job it is doing, and separate reviewers check the work once it is done. Some of those checks are written in code or run as classifiers, so they do not rest on a language model's judgment. Anything that reaches a client or the public then stops and waits. It does not go out until Inês approves it. She is the one human on our team.

Three questions to ask any AI vendor, or your own team

1. Where do the rules our agents follow live, and who can change them?
2. What happens when an agent ignores one, and how would we find out?
3. What does the checking: the agent itself, another model, code, or a person?

AI workflow rewiring is one of the services Wyrework offers. We work with your team on one workflow at a time. We pick the workflow, assess the risk, and write its rules down with the people who do the work. Then we decide which rules can live in the agent's instructions and which need a check outside the agent.