The Promise None of Them Kept

Promise theory rests on three axioms: an agent is autonomous and cannot be coerced, an agent can only promise its own behavior, and an agent’s knowledge of the world is local. From these axioms a control loop emerges where the agent first observes its environment, then reasons locally about what it saw, and finally makes a voluntary commitment to act. This loop does not rely on a central controller or external state snapshots; the agent’s promise is assessed only by whoever observes it. The theory emphasizes self-regulation rather than imposed commands.

Terraform’s plan/apply cycle operates on two files: the configuration and a stored state file. It diffs these files to generate API calls, which means it never observes the live infrastructure directly. Between runs Terraform is blind unless a human invokes refresh or apply, so it fails the observation axiom. The reasoning step is merely a comparison of two data structures, with the logic pre-computed by the human author, violating the local reasoning requirement. Finally, Terraform imposes changes on remote resources without promising anything about its own ongoing behavior, breaking the voluntary commitment axiom.

Ansible’s default mode pushes tasks from a control node to target nodes over SSH. The target nodes run a setup module to gather facts, but they have no authority to decide whether to act; they simply execute what the control node sends. When the control node is unavailable or connectivity drops, convergence stops and drift can go undetected, showing a lack of autonomous observation. The reasoning is performed on the control node, not locally on the target, so the local reasoning axiom is not satisfied. Moreover, Ansible imposes behavior on targets without those targets promising to accept it, which contradicts the voluntary commitment principle.

The article notes that large language models now supply the missing reasoning layer that promise theory always assumed. An LLM-enabled agent can observe a live system, weigh its findings against unenumerated context, and decide whether to act, fulfilling the observation and reasoning axioms. A prototype called swamp is described as being built for such an agent, suggesting a path toward infrastructure tools that can autonomously keep their promises. This shift would require engineers to rethink trust, safety, and the operational costs of deploying self-governing agents.