A code review helps detect defects. It also helps people understand why a solution was chosen, discuss an alternative and discover a part of the product they know less well.
When production and some of the review move to agents, these opportunities to learn change. Work can progress without people working through the questions that made it instructive.
This question came up at the Dev With AI conference. It extends beyond developers: a summary, a mock-up or a recommendation can also be produced without its recipient having developed the understanding needed to assess it. How do we keep learning under these conditions?
Completing a task and learning to master it
The study published by Anthropic in January 2026 offers limited but useful insight. It examines learning a Python library with or without AI assistance. The assisted group scored lower on average in an immediate assessment of understanding. The study does not support conclusions about long-term learning or all uses of agents.
Interactions focused on understanding were associated with better results, without establishing a causal effect for each practice.
For a team, this suggests looking at two outcomes: what has been produced and what people are now able to understand or do. Both can improve together, but the second needs room of its own.
Choose what people still need to understand
Preserving every former task because it used to teach something would be of little value. Some repetitive operations can be delegated without losing an important skill. Others bring people into contact with decisions the team will still need to know how to make.
The distinction depends on the work. For a technical change, understanding the data affected and possible effects on other journeys may be essential. For product analysis, people need to connect conclusions to observations and recognise what remains uncertain.
We can ask the person responsible for the work to explain their choice before presenting the agent’s response. What options did they consider? What would change their mind? What result do they expect to observe? These questions make their understanding visible without requiring them to redo the entire task.
Use AI to explore a question more deeply
An assistant can help people learn when asked to compare two options, explain a dependency or propose a case that challenges a solution. The conversation becomes useful if the person can then test the explanation against the product, a reliable source or an experiment.
Consider one possible exercise: before carrying out a change, a developer predicts which behaviours should stay the same and which should change. They then examine the results with a colleague. The differences provide material for understanding the system.
The same approach applies to analysing user feedback. Someone formulates their interpretation, compares it with the AI’s and returns to the sources to resolve disagreements. The aim is to make the reasoning open to discussion and verification.
These practices are possible ways to organise learning. Their usefulness needs to be examined within the team, taking its experience and constraints into account.
Keep people talking about decisions
A fully delegated review can remove the conversation between two people. We can preserve that conversation around choices worth understanding: a business rule, an architectural trade-off or a scope decision.
This does not require bringing the whole team together for every change. A focused discussion may be enough when people arrive with the proposal, its rationale and the checks carried out. Part of a senior colleague’s role is then to make their criteria visible, including what concerns them and what they are willing to leave unresolved.
Incidents offer another opportunity. Understanding why a check let a defect through can help both the team and its agentic software factory improve. If the agent silently fixes the issue and closes it, an opportunity to learn may disappear. A brief discussion of the cause and the change made to the check preserves that knowledge.
Make time for development
If a team’s objectives focus solely on delivery volume, activities that build understanding may get pushed aside. Managers can instead make the skills to develop explicit and create opportunities to exercise them.
For someone less experienced, this may involve bounded tasks, accessible support and gradually increasing responsibility. For a senior colleague, the challenge may be learning to design checks and organise delegation without becoming the approval point for every piece of work.
Coaching helps people examine situations where they hesitate, delegate too much or take everything back on themselves. Training provides space to practise particular skills. Everyday work then needs to allow people to maintain them.
At The Spark, we connect training, coaching and team support with this capacity to decide and act. AI provides new ways to produce work; the organisation also needs to help people develop the judgement that work depends on.