The method
From classic profile to an agentic profile & curriculum.
A five-step pipeline with a human checkpoint at every consequential step. The method is discipline-agnostic; ICT is the worked example throughout.
The centre of gravity moves from making to judging.
In an agentic world the work of a junior professional shifts from producing to specifying, orchestrating, verifying and being accountable for work that agents largely carry out. The lever moves up: not typing faster than the agent, but being the human-in-the-loop who decomposes problems, directs agents, judges their output and owns the result.
The counter-intuitive consequence: fundamentals matter more, not less — you can only verify what you understand. So a curriculum should not “add AI as a course”; it should move the centre of gravity from making to judging.
Surface
Ingest the current profile + curriculum; decompose the role into concrete tasks.
Classify
Position each task against the agentic-AI reality and label it.
Synthesise
Derive the new profile: workflow · tasks · capabilities · mindset.
Map
Translate capabilities into learning outcomes and modules; run the gap analysis.
Validate
Multi-model critique, framework conformance and field evidence.
The Curriculum Designer implements this pipeline. From intake and field-of-practice analysis through curriculum creation to the matrix and final report — with a human gate at every stage. The assessment workspace supplies the role analysis that feeds it.
For every task in the decomposed role, Nascith assigns a class (what happens to the task) and a direction (the verdict per capability). The same four classes and four directions apply to any profession.
Automate
Agents perform it end-to-end with light oversight. The task leaves human scope; the capability becomes supervision.
Augment
Human + agent together; human judgement elevated. New sub-skills: prompting, verification.
Emergent
Did not meaningfully exist before. New tasks: orchestration, context engineering, agent evaluation.
Durable
Resists automation — the human core. Retained, and often emphasised more strongly.
Needed more than before — e.g. critical evaluation of AI output.
Needed less — e.g. manual rote production.
Still needed but reshaped — “writing code” → “specifying, reviewing, integrating agent-written code”.
Did not exist — e.g. designing human–agent task hand-offs.
“Stroomwerk B.V.” — 2028
A feature no longer starts with “who picks up this ticket to code it” but with “how do we decompose this into tasks a coding agent can reliably perform, and how do we verify that?” A four-person team delivers what twelve once did. What Stroomwerk pays a junior to do:
- Decompose problems into agent-executable tasks with clear acceptance criteria.
- Specify instead of type — precise prompts/specs, context, examples and limits for agents.
- Verify at scale — review agent output for correctness, security and maintainability.
- Orchestrate — wire multiple agents/tools into a working, observable system.
- Guard architecture, data & integration — the boundaries you do not hand to an agent.
- Evaluate as a discipline — test agent behaviour: evals, regression, hallucination/injection checks.
- Be accountable — explain why it was built this way, ensure compliance, answer for incidents.
Stroomwerk would rather hire someone who codes less but can judge agent output razor-sharp, than someone who codes fast but misses an SQL injection the agent introduced. Judging requires deeper fundamentals than writing. Augmentation raises the bar for people; it does not lower it.
| Task | Class | Direction | What it becomes |
|---|---|---|---|
| Write CRUD / boilerplate | Automate | Less | Recognise & review generated code, not hand-produce it. |
| Design system architecture | Augment | Different | More synthesis, less typing; guard interfaces & long-term maintainability. |
| CS fundamentals, data structures & algorithms | Durable | More | The anchor for verification — you can't judge what you don't understand. |
| Security thinking | Durable | More | Agent code & agent systems open new attack surfaces; review is core. |
| Agent orchestration & tool/MCP design | Emergent | New | Multi-agent pipelines, tools/MCP servers, vector stores. |
| Spec- & context-engineering | Emergent | New | The new “programming”: precise instructions, context, examples, limits. |
| Eval design for non-deterministic systems | Emergent | New | Golden sets, regression on agent behaviour. |
| AI governance & EU AI Act | Emergent | New | Risk classification, transparency, audit trail — by design. |
| Verify & take accountability for agent output | Emergent | New | Provenance, explainability, answerability to client & regulator. |
| Negotiate requirements with a client | Durable | More | Human-judgement confidence; the relationship stays human. |
A · Fundamentals & Verification
The durable core, reframed as “being able to judge what an agent makes”. CS fundamentals, data structures/algorithms, architecture, security, debugging. Assessed without an agent — where the student proves the understanding is real.
B · Agentic Engineering
Light in year 1 → heavy in years 3–4. Spec/context-engineering, agent workflows, tool/MCP design, vector stores, multi-agent architecture, observability/FinOps, agent security, agentic UX.
C · Responsible & Governance
EU AI Act & GDPR, risk classification, data sovereignty, transparency & audit trail, ethics of human–agent teaming, accountability. Built into every project, not a separate ethics course.
D · Profession & Human–Agent Teaming
The human as director: delegate and stay accountable, explain agent output to client and regulator, work in teams where agents are members.
A working product no longer says anything about its maker — the agent may have built it.
- Review-viva — defend the flaws, risks and choices in agent-generated code, live.
- Eval-portfolio — ship the evals that prove you covered the behaviour, not just a solution.
- Spec-to-system — graded on specification and orchestration, not typed lines.
- Accountability-defence — “explain why this meets the EU AI Act and what you do at an incident.”
- Fundamentals without an agent — one deliberate line where understanding is shown unaided.
Run the method on your track.
The same pipeline, a different fictional company, a different delta — for any institute. Request the full specification or start a conversation.