A sandbox for one person
- Campus-licensed account, not a personal one
- Public or synthetic data only
- No production credentials
- Throw away or promote
8.1 — Capacity and pathways
When building is cheap, the hard part is people and judgment: who writes the spec, who checks the output, who owns the result. Here is what it takes to move an idea from one person’s experiment to something others rely on, with gates scaled to the risk.
Competencies
Effort moves from writing code to deciding what to build and checking it. This builds on the skills table; the right-hand column says where each skill matters most.
| Competency | What it looks like | Matters most for |
|---|---|---|
| Writing specs and acceptance criteria | What it does, what it doesn’t, how you’ll know it works, written before the agent starts | Functional analysts, enterprise teams, platform makers |
| Designing verification when not every line is read | Tests from the spec, protected test files, and diff review for what was removed as well as added. Spec-driven | Enterprise teams, research software engineers |
| Scoping agents | Permissions, context, budgets and stop conditions set before a run | Platform makers, IT, anyone running agents |
| Reviewing AI output critically | If you can’t explain it, don’t ship it | Everyone; faculty teaching it |
| Knowing the data | Its classification, authoritative source and steward | Business users, analysts, researchers |
| Knowing when not to use an agent | High stakes, unfamiliar systems, regulated data or no way to check the result | Everyone |
Environments
Three tracks, each with its own data rules. Moving between them is the promotion step.
People use personal accounts when the approved path is slower. Licensed access granted in days, starter templates, provided hosting and a review that is quicker than working around it do more than a policy can. Over-governing has its own cost; see proportionality and the shadow tool case.
Patterns
None of these needs a new office. Each puts domain knowledge and technical judgment in the same room.
The person who knows the work builds; IT or a platform team supplies the environment, credentials, integration and review.
Builders across units share specs, prompts, guardrails and mistakes. It also shows where tools overlap before they multiply.
A second person reviews diffs, tests and data access before others use the tool, scaled to the tier. A research software engineer plays this role for research code.
On campus
Emerson’s shared environment and doctrine
At Emerson, the people building knew the operations; none needed to be programmers. IT leadership supplied a shared approach: a written development plan, the “doctrine,” fed into every AI session, with architecture decisions recorded back into it, and a schema designed before features.
That combination let seven systems reach production. Read the case
Pathway
Each step adds the least structure that’s still safe. Most experiments should end as experiments; the gates are for the ones others come to rely on.
| Experiment | Pilot | Production | |
|---|---|---|---|
| Who uses it | The builder | A defined group, for a set time | Others depend on it |
| Data | Public or synthetic | Internal; confidential with steward approval | As classified, in an approved environment |
| Control tier | T1–T2 | T2–T3 | T3–T5, set by data and audience |
| Validation | Does it do the job? | Acceptance criteria tested with real users | Tests from the spec in CI; accessibility testing; protected test files |
| Security for AI-generated code | No secrets in code; no production credentials | Dependency and secret scanning; scoped credentials | Static analysis, dependency checks, human review of every diff including removals; security review at tier |
| Data protection | Nothing leaves the sandbox | Minimum necessary data; access logged | Retention, role-based access and audit logs outside the builder’s control |
| Ownership | The builder | Named owner, review date, registry entry | Named owner and backup, support path, offboarding plan |
| Gate to the next step | Someone else wants to use it | Pilot measures met and an owner accepts it | Reviewed against the tier on a schedule |
The gates draw on the promotion signals and quick test, the control tiers, shared responsibility and the governance checklist.