What is a sandbox workspace?

A sandbox workspace is a separate environment, set up like the real one, where people can work with AI agents on practice or test cases without any real customer, payment or record being affected. It is where reviewers learn to read verdicts and approve held actions before their decisions count.

Why does a sandbox workspace matter?

Nobody should approve their first held payment on a live queue. A sandbox lets people make the mistakes that teach most, such as approving a draft backed only by an external document, and see the consequence without harm. The agent behaves as it would in production; the difference is that nothing leaves the sandbox.

It also lets a team rehearse rare events on demand. A withheld action, a lapsed certificate that drops an agent to depth 1, a refusal because a request fell outside role scope: these may not occur on a live queue for months. In a sandbox, a trainer can stage them in an afternoon.

How does it work in practice?

A useful sandbox has four properties:

1. Faithful: the same screens, verdicts and record as the live workspace, so habits transfer.
2. Isolated: no connection that can move money, send messages or change real records.
3. Seeded: loaded with realistic, anonymised or invented cases that include the hard ones.
4. Reset: easy to restore to a clean state so each learner starts fresh.

Trainers then feed it cases. Many teams run a simulated queue inside the sandbox, so learners face a stream of mixed work at a realistic pace.

An example

An insurer onboards a new underwriter to an agent that drafts quote summaries and holds any quote above a set sum insured. In the first week, the underwriter works only in a sandbox seeded with forty past cases, five of which contain a broker's declared figure that conflicts with the policy library. She approves two of those five before spotting the pattern. The trainer walks through the record with her. By the end of the week she catches all five on a fresh set. The guide to onboarding an underwriter to an agent workspace describes the full sequence; there is a parallel guide for onboarding a paralegal.

How is a sandbox workspace different from a tabletop exercise?

A sandbox is a place where people act: they click, approve, send back, and the agent responds. A tabletop exercise is a discussion around a scenario, with no system involved. A sandbox trains hands and habits; a tabletop trains coordination and judgement about who does what.

A sandbox also differs from a test environment used by builders. Builders test the agent; learners in a sandbox test themselves. The same environment can serve both, but the cases and the measures are different.

Which terms are related?

• Simulated queue: a practice stream of work, usually run inside a sandbox.
• Lab exercise: a guided task with a set outcome.
• Shadow period: practice on the live queue alongside an experienced reviewer.

Senior leaders benefit too; see training senior leaders without wasting their time. For the full approach, read training teams for AI agents.

Where the Academy fits

Every lab in Pryme Intelligence Academy runs in a free Build workspace, so learners practise on the real product rather than a mock-up. The verdicts, the decision chain and the record they meet are the ones they will see on a live workspace.