Skip to content

B.1 Getting Started

Choose a recurring unit of work that an agent can usefully perform and whose result you can inspect. A service change, a document transformation, or a deployment procedure is a more useful starting point than an organization-wide commitment to autonomy. Name the artifact being changed, the expected outcome, the point at which it can affect someone else, and the engineer responsible for accepting that consequence. A bounded first attempt makes both failure and improvement easier to interpret.

Before building machinery, watch how the work is done now. What must someone reconstruct from source, documents, or conversations? Which decisions recur? What must remain true while the implementation changes? Preserve the requirements and decisions that matter to acceptance, including qualitative obligations that still require judgment. Also identify choices that can safely remain open. A brief that specifies every implementation decision can consume the very engineering effort delegation was meant to save.

Inventory the knowledge and enforcement you already have. A schema may reject invalid data, and a permission may prevent a dangerous action. A test may expose a regression, while a review may settle an ambiguous case. Determine which of these actually affects whether work proceeds. An adequate existing mechanism needs a place in the workflow, not a replacement bearing a new name. Where the obligation is already enforced economically, additional system modeling may contribute little. Where people repeatedly recover the same architectural relation, a small representation may help before any new gate is needed.

Use a Change as the First Model

A structured ticket can preserve desired behavior, rationale, acceptance criteria, affected concepts, constraints, related changes, and prior decisions. When it carries enough information to reason about one requested change, it functions as a purposeful change-scoped model. It should connect the change to longer-lived system knowledge rather than repeat that knowledge in a growing collection of tickets. For a defect, record the observed discrepancy and the intended correction; diagnosis may still be the difficult part of the work.

Give those connections a consumer. An agent may follow the ticket’s component reference to the relevant ownership model; a reviewer may follow an invariant to its tests; a completion audit may compare the requested behavior with the landed result. Add a field when someone or something will use it. Fields without a consumer become additional material to keep current. The first useful investment may be a reliable link between an existing ticket, an architectural decision, and its acceptance evidence.

The ticket’s lifetime matters. A ticket written in 2022 need not describe the product in 2026, but it can remain an accurate representation of the intent governing the 2022 change. Its persistence is useful history rather than a synchronization burden. A current architectural model makes a different claim and must be reconciled when the relevant system or intent changes. Mark that distinction before installing freshness checks: otherwise a well-intended audit can classify valid history as drift.

Establish the Delegation Boundary

For the first trial, state what the agent may change, what evidence will support acceptance, and what should stop the work. Separate the producing agent’s completion report from the evidence used to evaluate it. Existing tests, analyses, measurements, and review procedures may already provide much of that evidence. When they do not settle a consequential question, name the remaining judgment and who must exercise it. Keep an escalation path for a mistaken premise, an unavailable dependency, or an obligation the environment cannot adequately evaluate.

Run a small number of comparable work units and retain the interventions they required. Distinguish time spent creating new intent from time spent reconstructing known facts, checking routine conditions, or recovering from a recurring failure. These observations tell you what the next investment should address. A repeated escalation can indicate an appropriate human judgment boundary; it does not automatically justify a new evaluator or permission.

Expand the scope or consequence of delegated work when the evidence supports doing so. Check that someone remains able to inspect the result, diagnose failures, and continue essential work when the preferred agent or service is unavailable. Review the boundary by engineering surface; a successful workflow does not establish that the rest of the product is equally governable. This keeps the adoption decision connected to demonstrated capability and its costs.

Then stop. Once the selected capability has adequate representations, evidence, and enforcement, further machinery needs a new reason to exist. Stop when additional representation costs more to maintain than the reasoning it saves. A useful first result can be modest: one dependency that stops being rediscovered, one recurring review decision made explicit, or one acceptance condition that no longer depends on memory.

© James C. Davis, 2026–present