The Groundwork

The Groundwork is the four-document protocol that every AI agent (Post) must complete before it runs in production. The four documents are: scope.md (what is in and out), contracts.md (behavioral rules and constraints), testing.md (how you verify it works), and prompting.md (the actual instructions the agent receives).

This is not a template to complete once and file away. These are living governance documents that define the operating contract between the organisation and the agent. When the agent drifts, you trace back to The Groundwork. When the agent is audited, you audit The Groundwork. When the agent is improved, The Groundwork is what changes.

The protocol exists because the cost of deploying an ungoverned AI agent is not the deployment itself — it is the compounding cost of every decision that agent makes without boundaries, every output it produces without testing criteria, and every failure it encounters without an escalation path.

scope.md defines the territory. What this agent handles, what it never handles, and the explicit boundary between the two. The most important section is "out of scope" — explicit exclusions prevent the agent from gradually expanding into territory it was never designed for.

contracts.md defines the frozen interface. Any output format the Post must produce, any behavior that cannot change without Station Owner authorization, any schema that downstream systems depend on. A behavior change without a contracts.md update is a breach. That framing forces the question: did we authorize this change or did the Post drift?

testing.md defines the golden vector protocol. Deterministic input-output pairs registered before the Post is trusted to run unsupervised. The gates are binary: pass or blocked. No partial pass, no almost-correct. The testing gate exists because "does this Post work" must be answered with evidence, not confidence.

prompting.md is the institutional context layer. Everything the Post needs to know about the organisation, the Station, the people it serves, and the operating reality it is embedded in. This is not a prompt template. It is the persistent context that makes the Post coherent across interactions. A Post without this layer is not governed — it is just competent in isolation.

A Post hallucinated. It told a customer something that was not true. The information was plausible, stated confidently, and wrong. The instinct is to blame the model. That framing leads to one set of solutions: get a better model, write a better prompt, add a disclaimer. None of those close the governance gap.

The actual cause, almost always, is one or more Groundwork failures. The Post had no golden vectors in testing.md that covered this type of query. The Post had no output contract in contracts.md defining what a correct response looks like. The Post had no scope boundary in scope.md that would have triggered a Relay when the query exceeded what it was built to handle. The Post had no institutional context in prompting.md that would have given it the accurate information it needed.

Every one of those is a Groundwork failure, not a model failure. The model did what models do when they encounter a query with no governance: they synthesize something plausible and present it confidently. That is the default behavior of an ungoverned agent. A governed Post with an average model will outperform an ungoverned Post with an excellent model every time, because governance is what makes behavior predictable.

The most common objection is speed. A startup moving at breakneck pace will say: I do not have three days to write documentation for a scraping agent. I will just write a prompt and fix it when it breaks.

This works until it does not. The first failure is forgiven. The second is debugged. By the third, the team has accumulated enough ungoverned automation that the debugging time exceeds the time The Groundwork would have taken.

Waypost accounts for this with Provisional Commissioning. You can launch a Post with only scope.md and a Directive. That is the minimum viable Groundwork. The Post runs in Pressure State permanently until the remaining documents are completed. Every run generates an entry in the Debt Ledger. Every output requires human review before it ships. No gate gets removed until the full Groundwork is in place.

The Groundwork Compiler at waypost.run generates all four documents in about ten minutes based on three questions. That output is a starting point — about sixty percent complete — and requires another hour or two of specific work to reach Standard State quality. That is a very reasonable front-load for the operational coverage you get in return.