Transforming Software Development one Agent at a Time
Posted on: 2026-03-02
My journey into professionally developing agents that perform software engineering tasks is going beyond having the system simply write code. As I work more and more with agents, I realize that they are excellent at following very clear directives, but they often won’t think further. A part of me sees interesting behavioral similarities to when my children were younger. If you ask them to pick up their cars, they will, but they will leave the race tracks on the ground. Similarly, if you ask an agent to update a value threshold, it might change the default value of a function but ignore the configuration file.
While I cannot disclose too many details of the actual system I am building, I can share that I am borrowing more and more concepts that humans have been using for a while. The sense of responsibility and authority found in a strict structure, like in the army, helps. Dividing to conquer works splendidly to create small units of work while optimizing context. Having specialized agents helps with tweaking and improving each part of the system, much like having specific experts in a team. Similarly, having planners and people who verify the quality of others' work is beneficial in an AI environment. After weeks of work, I have a small militia of agents slowly becoming an army of increasingly powerful elements.
In this journey of creating a self-guided army of builders, the landscape of software engineering shifts from writing lines of code to improving how they are generated. The life cycle changes as more people participate in the creation of features, with engineers fixing the issues that their own systems generate. In the long term, this transformation of many agents building code might converge toward a few vendors hosting these solutions alongside integrated code repositories and deployment infrastructure. For now, on-premises solutions remain the safest and most hybrid options to hook into existing workflows, and they are indeed the most cost-effective when relying on internal ML systems.