On 28 October, GitHub outlined its direction for the next generation of software development: multiple AI agents working across multiple repositories, with people directing and reviewing their work. The ambition is to automate more of the work between idea and delivery. The direction is clear. But a product promise is no proof that autonomous development already works at scale.
This deserves a more informed discussion than yet another prediction about how many developers AI can replace.
More code is a poor measure of progress
When agents can perform more development tasks, the bottleneck shifts. The ability to define the problem, set boundaries and assess the result becomes more critical as producing the code itself takes less effort.
An agent can implement a change. Deciding whether that change is right for the business takes more. A passing test does not automatically tell you whether the architecture is sound, access permissions are appropriate or the solution can be maintained.
The argument is simple: as execution becomes cheaper, judgment becomes more valuable.
That is a reason to use automation with greater care and precision. Otherwise, we risk getting better at generating activity and worse at recognising actual progress.
Who understands the code when the agent is done?
There is also a question of expertise. Researchers at the University of Gothenburg describe how AI is changing the way developers share knowledge. Questions that previously went to colleagues now increasingly go to AI. They also highlight the risk that faster problem-solving does not always lead to deeper understanding. University of Gothenburg
This poses an uncomfortable question for development organisations: if we automate away parts of the work through which understanding develops, how do we maintain the expertise needed to review the result?
Having a person at the approval button is not enough. Review needs substance. Someone must understand the wider system, be able to challenge the agent’s solution and have the authority to stop a change.
Automate the work. Own the outcome.
For technology leaders, choosing which AI tool their teams should use is therefore only part of the job. They also need to decide which tasks agents may perform, which decisions require human judgment and what evidence is needed before a change can reach production.
This is where software development, AI and cybersecurity come together around a concrete task: turning automated work into reliable systems.
Coding can be delegated. Accountability stays with us. The more independently AI works, the more important people’s ability to review its work—and their authority to stop it—becomes.
More AI in development demands better judgment. We help you bring software development, AI and cybersecurity together—from clear boundaries and authority to review before production.
Talk to us about your AI-driven development.
