A capstone project is a substantial final piece of work in which a learner applies everything a course taught to one realistic problem, and is judged on the result against set criteria. It closes a programme by asking the learner to produce something, not to recall something. For AI agents, the most telling capstone is an agent that passes certification.
Why does a capstone matter?
Builders and architects make many linked choices: what the agent's brief says, which sources it trusts, how far each capability may act, what is held for a person. A test of each choice in isolation does not show whether they fit together. A capstone does. If the brief is vague, the agent fails its grounding gate; if the bounds are loose, it fails on scope.
It is also strong evidence for others. A regulator or client asking for proof of competence gets a working artefact and a record of how it was judged, which goes further than a pass mark. That is part of how to show AI literacy to a regulator.
How does a capstone work in practice?
A capstone for agent work usually has four stages:
- Scope: the learner picks a real job, such as drafting replies to supplier queries, and defines what the agent may and may not do.
- Build: they write the brief, connect sources, set bounds and decide which actions are held for a named person.
- Certify: the agent is replayed against real cases and scored on each gate. It either clears or it does not.
- Defend: the learner explains their choices, and in senior tracks a panel questions them.
The outcome is objective in one part (did the agent clear certification?) and judged in another (are the design choices sound?).
What is an example?
A learner builds an agent to draft responses to invoice queries for a finance team. At first certification it fails the refusal gate: it answers questions about payment dates for suppliers whose records it cannot see. The learner narrows its sources, adds a rule to hold any answer that relies on an external document, and recertifies. It clears at depth 2: the agent drafts and a person sends. The capstone shows not just that the learner can build an agent, but that they can read a failure and fix it.
How is a capstone different from a practicum?
A practicum is a period of supervised real work, judged across many decisions over time, such as two weeks of reviewing an agent's held actions. A capstone is one built deliverable, judged as a whole. Operators prove themselves through practicums; builders and architects through capstones. Both are stronger than a course alone, a point made in whether an AI literacy course is enough for staff who approve agent actions.
Which terms are related?
- Board review: a panel examination of the capstone in senior tracks.
- Attestation of training: a weaker form of evidence, explained in what an attestation of training is.
- Checkable credential: the certificate that records the result, discussed in checkable credentials as a hiring signal.
The wider topic is AI literacy and certification for people who work with agents.
Where the Academy fits
In Pryme Intelligence Academy, Certified Pryme Builder requires a proctored exam and a capstone agent certified in Attest to depth 2 or higher. Certified Pryme Architect adds a capstone and a board review with two Pryme assessors. Every lab, including capstone work, runs in a free Build workspace.
Questions people ask
Can a capstone use real company data?
It can, if the organisation approves and its data rules allow it. Many learners use a sample or test copy of real cases so the capstone reflects their actual work without exposing live data.
What if the capstone agent fails certification?
That is information, not the end. The learner reads which gate failed, changes the brief, sources or bounds, and runs it again.