A summary of the feedback on the week 2 [AI] exercise (First contact with a coding agent).

39 respondents.

Background questions

Previous experience with coding agents

Experience Respondents Share
No experience 23 59 %
I have tried them sometimes 10 26 %
I have used them to some extent 3 8 %
I use them almost daily 3 8 %

For most respondents, the exercise was their first contact with agents.

Agent used on the course

Agent Respondents Share
GitHub Copilot 25 64 %
OpenCode 14 36 %

Comments on the week 2 exercise (28 responses)

The overall tone is positive: the exercise was considered a good and instructive introduction to agents. Many, however, found it light and short.

Recurring themes:

  • Limited tokens were clearly the most common complaint. Trial and error felt stressful, and tokens ran out fast even with small models such as GPT 5.5 mini. Respondents also wondered why even the “unlimited” models consume the limited quota.
  • Problems with Copilot: Copilot felt confusing, its changes were hard to follow and the inline feature misbehaved. One respondent could not get Copilot to work at all and skipped the exercise.
  • OpenCode was praised. Respondents appreciated that it was quickly added alongside Copilot, and working on the command line got people excited. On Windows, however, the configuration took some struggling.
  • The benefit for small changes was questioned. According to one respondent, small incremental changes are faster to make without an agent, because the biggest time sink is the “overhead” of using the agent rather than the waiting.
  • Code quality and understandability: The generated code worked, but its style varied. Code written by an agent is harder to check than your own, so errors can easily slip through. Some edited the code themselves to keep their understanding of it.

Opinions on the spread of AI (31 responses)

The stance is mostly pragmatic and cautious: agents are seen as inevitable and useful, but they require supervision and understanding.

  • Positive: Agents speed up mechanical and repetitive tasks such as tests and lint fixes. They also make it possible to work at a higher level of abstraction, and skills in using them are valued in working life.
  • Understanding as a condition: A recurring thought is that a human must understand what the code does and ultimately approve the changes themselves. Vibe coding does not replace software engineering skills.
  • Concerns:
    • the job market and especially the position of beginners (“like having bought Teboil in 2022”)
    • growing demands for efficiency
    • erosion of one’s own skills and weakening knowledge of the codebase
    • “slop” and settling for bad code
    • climate impact
    • impostor feeling when the work does not feel like your own
  • The sharpest comments: “scary”, “unfortunate”, “has brought nothing but sorrow and disgust”.
  • The study perspective: One respondent feels that the workload of their studies forces them to use AI, because there is no time left to debug errors the traditional way.

Areas for improvement based on the feedback

  1. Token quotas and how they work should be explained more clearly, or the quota could be increased.
  2. Setting up Copilot needs better support. OpenCode seems to work more smoothly.
  3. The exercise could be more challenging or larger if the benefits of agents are to be shown. With small changes, the benefit of an agent easily remains invisible.
  4. Concerns about working life and one’s own learning should be addressed on the course, for example through working-life perspectives, since students asked about them directly.