Should HBO-ICT Students Be Allowed to Use Claude Code and Codex?
A nuanced framework for deciding when agentic coding tools make educational sense in HBO-ICT, based on learning outcomes, student level, task type, evidence, and professional relevance.
TraceYield Insights
11 min read · Updated September 15, 2026
TraceYield
AI engineering evidence
Should HBO-ICT Students Be Allowed to Use Claude Code and Codex?
The answer depends on what the course is teaching
“Should students be allowed to use Claude Code or Codex?” is not one question. It can mean whether a first-year student should use an agent to learn syntax, whether a project team should supervise repository changes, or whether a professional-practice course should evaluate responsible tool use. Each case has a different educational rationale.
A useful decision starts with the learning outcome, not the popularity of the tool or fear of its capability.
Consider student level and task type
Early learners may need controlled experiences without agents to build mental models and debugging habits. They may later use assistance for explanations or constrained comparisons. Advanced students can work on authentic projects where agents are part of the professional environment, provided they can verify and explain the result.
Task type matters too. A closed exercise intended to assess independent fluency differs from an open project intended to assess architecture, integration, and responsible development.
Agent capability changes the evidence requirement
A tool that can read project files, edit multiple files, run commands, or iterate on tests creates a different assessment context from inline autocomplete. Course teams should describe the capability they are permitting rather than rely only on a brand name. The same vendor can expose different modes and permissions.
Where agents are allowed, evidence should include material use, changes, verification, and explanation. Students should know what remains their responsibility.
Permitted use can be educationally valuable
Students can learn to specify context, evaluate alternatives, identify hallucinations, review generated changes, and make responsible trade-offs. Those are relevant professional capabilities. Allowing a tool can also make the course more authentic when graduates will encounter similar systems in practice.
But access alone does not teach those capabilities. The assignment and feedback must create a reason to question and verify output.
There should still be controlled work without agents
A programme needs some evidence of individual knowledge and skill that is not outsourced to a tool. That may be a supervised practical, live debugging, oral explanation, or a transfer task. Its form should follow the outcome and support fairness rather than act only as punishment for permitted use elsewhere.
A balanced approach can combine an AI-permitted professional lane with a no-agent learning lane. The two lanes answer different questions.
A decision checklist for programmes
Before allowing a specific agent, ask: what outcome is being assessed; what can the tool do in the selected mode; what evidence remains; what data may be entered; what access route is fair; what disclosure is required; what verification is expected; and how will the course handle ambiguity?
The conclusion should be specific: allowed for this task under these conditions, limited to this purpose, or not appropriate for this outcome. “Yes” and “no” are both too blunt without the educational context.
A decision matrix
For foundational outcomes and high need for unaided fluency, use a no-agent or tightly limited lane. For evaluation, debugging, integration, and professional practice, an agent-permitted lane may be appropriate. For sensitive data or unclear evidence, use a controlled environment or a different task. For advanced projects, combine permission with disclosure and verification.
The matrix makes the decision explainable to students, lecturers, and examination boards. It also makes it easier to change the policy when tool capability changes.
Pilot the educational use, not only the product
A course team should evaluate whether students learned more, verified better, or made stronger decisions—not merely whether the agent completed tasks. Collect examples, student feedback, assessment evidence, and lecturer workload. A tool can be technically impressive and still be a poor fit for a particular learning outcome.
The decision should remain open to revision.
Do not confuse professional relevance with permission
The fact that employers use coding agents does not mean every course outcome should be assessed with an agent. Professional relevance is one criterion. Learning stage, task design, evidence, access, privacy, and assessment validity matter as well.
Conversely, the fact that a tool can complete a task does not mean it has no place in education. Students may need guided practice in supervising exactly that kind of completion.
State the rationale to students
Students are more likely to understand a mixed policy when the course explains which capability each lane is intended to show. “No agent for this practical because we are assessing independent debugging” is clearer than “no AI because it is cheating.” “Agent permitted here because we are assessing professional review and verification” is clearer than “AI is allowed.”
The rationale also gives the programme a basis for reviewing whether its decision worked.
Decide at assessment level
The programme can permit different levels of assistance across a single course. A supervised practical may be no-agent, a design workshop may use an agent for comparison, and a project may permit agentic changes with process evidence. The policy becomes nuanced without becoming vague when each decision is tied to an outcome.
Record the rationale so later teams can revise it rather than starting the debate again from zero.
Agent completion does not always mean engineering completion.
References
Pilot Program
Understand the WHY behind your engineering AI usage.
TraceYield evaluates trajectory evidence instead of stopping at spend totals. Join the private pilot to review AI coding usage with engineering context, security controls, and developer trust.
Join the TraceYield private pilot