Where will senior engineers come from if AI does a junior engineer's work?

The decisive moment in a project comes before implementation starts: deciding which problem to solve, how to structure it and which trade-offs to accept. AI can accelerate building a solution. The organization still needs people capable of giving it direction.

If people starting out receive tasks already defined and a generated solution, where will they learn to design a project? This is where I place one of the main human-in-the-loop roles: human participation that defines intent, designs the solution and structures the work before delegating execution.

The problem I want to address is how to develop that judgment while taking full advantage of the speed AI provides.

Code review is being automated too

Anthropic describes a workflow that automates review and allows low-risk changes to merge without human approval, after its checks. It retains human involvement in other cases, including architecture. This is a vendor’s experience, under its own conditions. Claude Code team case study, October 2026.

My expectation is that, in the near term, delegating code review entirely to AI will become common. That is why I want developing expertise to include defining the problem, weighing alternatives and designing acceptance conditions from the start.

Being senior involves making decisions with incomplete information, understanding dependencies and taking responsibility for consequences. These capabilities also develop through participation in how a project is framed. Experienced professionals need to practice them when entering a new domain.

Bainbridge examined how automation can reduce practice of skills that remain necessary. Her industrial analysis provides a conceptual precedent for this question. Ironies of Automation, 1983.

What research contributes to understanding learning

Shen and Tamkin observed lower immediate comprehension with AI in a trial involving 52 participants learning a Python library, without a statistically significant improvement in task time. It is a preprint about a brief experience. Study, 2026.

Brynjolfsson, Li and Raymond found productivity improvements and evidence consistent with learning in customer support, including during tool outages. It is a study of one company, with noisy outage estimates. Article, 2025.

Bastani and colleagues observed in school mathematics that a tutor with safeguards mitigated much of the subsequent unaided performance harm associated with open-ended assistance. Applying this to work requires caution. PNAS, 2025.

These studies do not evaluate how architects develop or predict professional careers. My interpretation is that assisted production and capability development each deserve attention. The proposal below is my application to engineering work.

An example: design before generation

Imagine a booking service that must tolerate communication failures. Before implementation, the team needs to decide what repeating a request means, how to prevent duplicate bookings and which guarantees must hold when concurrent requests arrive.

A timeout can leave the client uncertain whether the server completed the operation. RFC 9110 sets conditions for automatically retrying non-idempotent operations. HTTP Semantics, §9.2.2.

I would involve the learner in those decisions: drawing the flow, separating responsibilities, making assumptions explicit and proposing acceptance criteria. AI can then help implement and review. If a constraint changes, that person should be able to explain which part of the design needs reconsideration.

The example is hypothetical and leaves the complete solution outside its scope. The learning opportunity lies in structuring the problem and justifying decisions before execution.

Three practices for developing that judgment

Collins, Brown and Newman’s cognitive apprenticeship offers a reference: make reasoning visible, guide practice and gradually remove support. Its application to this workflow is a proposed adaptation. Cognitive Apprenticeship, 1987.

AI can participate throughout. Human mentoring should help people understand decisions and progressively take responsibility for them. Acceptance of the result retains an explicit owner; sensitive practice requires an isolated environment.

I would reserve time to design together, discuss alternatives and explain why some are rejected. If that work always falls to the same experts, I would incorporate its transfer into capacity planning.

I would observe two outcomes: improvement in delivery and quality, and more people’s ability to frame a project or adapt its design weeks later. A good immediate explanation provides only a partial signal; local change over time, alone, does not establish causality.

My position

I want to accelerate execution and take advantage of everything AI can automate. I also insist that the organization pay attention to developing judgment, sharing knowledge and preparing the people who will lead its next projects.

I would involve junior engineers in design decisions from the outset, with mentoring and increasing responsibility. That is where I want to focus their development.

Faster execution strikes me as excellent news. Making speed the organization’s sole focus strikes me as a poor leadership decision.

References

• Shen, J. H. and Tamkin, A. (2026). How AI Impacts Skill Formation. Preprint.
• Brynjolfsson, E., Li, D. and Raymond, L. (2025). Generative AI at Work. QJE, 140(2).
• Bastani, H. and colleagues (2025). Generative AI without guardrails can harm learning. PNAS, 122(26).
• Bainbridge, L. (1983). Ironies of Automation. Automatica, 19(6).
• Collins, A., Brown, J. S. and Newman, S. E. (1987). Cognitive Apprenticeship. Technical Report 403.
• Fielding, R., Nottingham, M. and Reschke, J. (2022). RFC 9110, §9.2.2.
• Osmani, A., Lunt, M. W. and Wang, Q. (2026). How the Claude Code team does code review. Anthropic case study.