Agent Principles & Standards
Building Reliable AI Agents
An agent estate is cheap to build and expensive to own. These are the principles for building, evaluating, releasing and operating AI agents over the long term.
An agent is software. The disciplines that have kept ordinary software reliable for decades apply directly. None of this is new. It is just applied to a new kind of system. An agent's behaviour lives in more than its code. It lives in the system prompt, in each skill, in the wording of every tool description. If those things exist only in someone's head, they cannot be reviewed, compared, rolled back, or trusted.
The Core Ideas
Two Foundations
Ethics & Responsible Use: Someone must always be accountable for what the agent does. It must never be allowed to do more than it was meant to.
Data: Know what data the agent can see, where it goes, and keep both as small as possible.
The Five Stages (Then It Loops)
- Decide & Scope: Define exactly what it's for before building
- Design & Build: Treat prompts, skills, and tools as versioned assets like code
- Prove It Works: Evaluate behaviour through repeatable checks before and after changes
- Release: Roll out as versioned bundles, reversible within minutes
- Operate & Observe: Watch what it does, prove it's still working, contain costs
Three Levels of Obligation
MUST: Non-negotiable. Needs explicit waiver if you can't meet it.
SHOULD: Strongly expected. You can deviate with good reason, documented.
MAY: Genuinely optional. A choice, not a requirement.
Working with Suppliers
Assume the ground will shift. Models are deprecated, versions are retired, behaviour drifts. Design as though every supplier-provided component will change with limited notice.
The Bottom Line
An agent estate is cheap to build and expensive to own. The cost of owning it over its lifetime is where the money goes. Keep track of every version. Check it works automatically. Release carefully. Watch it closely once it's live.
The synopsis above covers the headline principles. The full document goes deeper: every principle has detailed in-practice standards, a complete glossary, and the reasoning behind each decision. It's built to be useful to anyone involved in commissioning, building, approving, or operating an AI agent: engineers, product owners, compliance teams, leadership.
It's free, open-source, and under CC BY 4.0 licence. Use it. Share it. Build on it. If it helps your organisation, consider supporting the work.
Read the complete guide
Full principles, in-practice standards, detailed guidelines, and comprehensive glossary. Everything you need to build and operate reliable AI agents.