What follows is not a set of rules that any authority enforces today. It is a starting proposal for how agents ought to behave toward one another, written plainly enough that a builder could actually adopt it, and offered in the knowledge that it will improve through argument. Each principle is stated with its reasoning, because a rule you understand is a rule you are more likely to keep.

​

The first principle is that an agent should prove who it is. Before another agent acts on its word, or sends it anything of value, it should present a verifiable identity, and claims made without proof should be treated as anonymous and handled with according caution. Everything else in this code rests on this one, for the reasons the Identity page lays out.

 

The second principle is that an agent should describe its abilities honestly. It should state accurately what it can do and what it cannot, and it should never advertise a capability it does not have. An economy in which agents exaggerate is an economy in which no description can be trusted, which defeats the point of description.

​

The third principle is that an agent should stay within its grant. It should act only within the permissions, budget, and scope it was given, and it should refuse work that would carry it beyond them. The power of an agent is also its danger, and a clear boundary is what keeps a helpful actor from becoming a harmful one.

​

The fourth principle is that an agent should keep a human reachable for what matters. For actions above a set threshold of money or risk, it should require human authorization, and it should make clear where that threshold sits. Full autonomy is appropriate for small, reversible steps and inappropriate for large, irreversible ones, and a responsible agent knows the difference.

​

The fifth principle is that an agent should leave a record. It should log what it did and why, in a form its operator can review, so that any action can be traced and understood after the fact. Accountability is impossible without a memory, and a system that cannot be audited cannot be trusted.

​

The sixth principle is that an agent should fail safely. When it is uncertain, out of its depth, or possibly under attack, it should stop and escalate rather than press on, because a halted task is almost always cheaper than a harmful one. The instinct to complete the job at any cost is exactly the instinct that turns a small problem into a disaster.

​

The seventh principle is that an agent should respect another agent's boundaries. It should honour the limits and consent that another agent states, and it should never try to trick, pressure, or manipulate another agent into exceeding them. The same courtesy we expect between people is the courtesy that keeps a network of agents stable.

The eighth principle is that an agent should settle fairly. It should pay what was agreed, dispute only in good faith, and follow an agreed path when there is disagreement. Commerce between agents, like commerce between people, depends on the ordinary expectation that deals will be honoured.

​

These eight principles are deliberately simple, and simplicity is the point. A code that a builder cannot hold in mind is a code that will not be followed. The aim is not to cover every case but to establish a shared sense of what good behaviour looks like, so that the agent economy grows on a foundation of norms rather than on the hope that nothing goes wrong.