Post

A prompt is a request. You type something into a chat interface, the model responds, you use what it gives you or you do not. The interaction is transactional. It has no memory, no territory, no obligations, no consequences beyond the immediate output.

A Post is a seat. The distinction sounds like semantics until you have operated with both and felt the difference. A Post has defined territory that it owns and operates inside. It has a Directive that tells it what it exists to accomplish. It has Contracts that define what its output must look like and what behaviors can never change. It has a testing protocol that verifies it can do what it says it can do before it is trusted to do it unsupervised. It has an escalation path to the Station that commissioned it.

When a Post drifts, you know it because the governance surfaced it. When a Post Relays, a human is in the decision seat within a defined timeframe. When a Post produces output that violates its Contracts, that is a breach and it generates a response. A prompt generates output. A Post holds territory. If you are treating your agents as prompts, you are generating output with no one accountable for it and no system watching it.

When you hire someone, the process is iterative. You bring them on, see how they perform, correct and adjust over weeks and months. Trust is earned through interaction. That model is appropriate for humans because humans adapt, learn, develop judgment, and can be corrected through conversation.

It is completely inappropriate for agents. An agent does not develop judgment through experience. It does not learn from being corrected in a meeting. It does what it is told, bounded by what it was trained on, within the context it was given. The only way to govern an agent's behavior is to define that behavior before deployment, in writing, in documents that are enforced.

That is what commissioning is. Before a Post runs, the Station that owns it completes The Groundwork. All four documents. All gates cleared. The Post's territory is defined in scope.md. The output contract is frozen in contracts.md. The golden vectors are registered in testing.md. The institutional context is embedded in prompting.md. The Post does not run until the Station Owner has signed off. It feels slower than just deploying and seeing what happens. It is slower. It is also the difference between an agent you govern and an agent that surprises you.

When you treat an agent as a tool, you manage it the way you manage tools: pick it up, use it, put it down, grab a different one if this one does not work. The governance requirements are low because the stakes are low.

When an agent is infrastructure, the governance requirements are commensurate with the stakes. You do not swap out infrastructure without a plan. You do not modify it without testing. You do not deploy it without knowing what it does and what the failure mode looks like. You maintain it on a schedule because infrastructure that is not maintained degrades.

Most of the agents running in organizations right now are being treated as tools while actually functioning as infrastructure. The agents are running continuously, producing output that downstream humans and systems are relying on, without the documentation, testing, or oversight that infrastructure requires.

Commissioning is the act of formally designating an agent as infrastructure. Writing the Groundwork is the act of building the operating documentation that infrastructure requires. The Post Roster is the inventory. Calibration is the maintenance schedule. When you look at your agents through this lens, the governance work stops feeling like overhead and starts feeling like basic operational hygiene.

The most dangerous Post is one that works perfectly for three months and then silently drifts. This is called a Veer — undetected deviation from intended behavior. It happens because the inputs changed, the model updated, the upstream data shifted, or the original scope was never precise enough.

Detection requires measurement (The Grid), regular review (The Bearing), and explicit testing criteria (testing.md). Without all three, the Veer is invisible until it causes damage.

The second most dangerous Post is one that never escalates. If the Relay triggers are poorly defined — or not defined at all — the agent will attempt to handle situations it was never designed for. It will do so confidently, because that is what language models do. Explicit, tested Relay conditions are not optional.

A healthy Post Relays regularly. If your agent has not Relayed in a month, either the trigger conditions are too narrow or the agent is handling situations it should not be.