Most AI agents today still operate on the infrastructure of the Web2 era.

They log in to accounts, call APIs, wait for approvals, and hand control back to humans when exceptions occur.

This model is still feasible for automation scripts or assistant-type applications, but when agents are expected to act independently, the problems become immediately apparent.

The original design intent of Web2 systems is to serve human users, while agent systems require a subject that supports autonomous action.

The real limitation is not in performance, but in the underlying assumptions.

Web2 infrastructure assumes by default that:

• Actions are initiated by humans

• Credentials are long-lasting and permissions are either/or

• Payments require interactive confirmation and can be revoked

• Trust is constrained by policies and processes, rather than enforced at execution time

However, agent behavior precisely breaks these premises.

Agents that operate continuously need clear permission boundaries, rely on deterministic settlement mechanisms, and must be able to directly prove what they have done, rather than explaining afterward.

Forcing Web2 infrastructure to be adapted for agent use will only compel systems to adopt fragile workaround solutions, reintroducing dependence on humans, undermining security, and completely dismantling autonomy at scale.

Kite addresses this fundamental issue by using primitives where agents are first-class objects.

On Kite, identity, permissions, constraints, payments, and execution are directly enforced at the protocol layer, rather than relying on APIs or policies for indirect constraints.

This enables agents to act autonomously within clear and controllable boundaries, natively complete transactions, and generate verifiable on-chain records for every action.