Here Be Monsters: Maintaining Software Systems in the Age of AI Agents
The Shift from Coding to System Ownership
Software engineering is undergoing a fundamental transformation where the primary challenge is no longer writing code, but maintaining systems generated by autonomous agents. As AI-driven workflows become standard, the traditional path of learning by writing every line by hand is becoming economically and practically unviable. Instead, successful engineers must cultivate a high-level model of their systems, knowing exactly when to dive into the details and when to trust the automation.
The Three Pillars of Modern Ownership
To manage complex systems without needing to manually inspect every line of code, engineers must rely on three core pillars: fundamentals, breadth, and just-in-time depth. These skills allow an operator to maintain control even when the underlying implementation is opaque.
1. Fundamentals
Understanding the underlying principles of computer science—such as data structures, networking, concurrency, and distributed systems—is essential for identifying "quiet failures." When an agent produces code that exhibits unexpected performance degradation, such as quadratic complexity, a solid grasp of theory allows an engineer to recognize the issue immediately, regardless of the specific implementation details.
2. Breadth
Breadth provides the "map" of the system. Successful engineers use their knowledge of disparate disciplines to weave together solutions and identify which parts of a system are prone to risk. This high-level perspective acts as a guide, helping engineers determine where to focus their attention and which areas require deeper investigation.
3. Just-in-Time Depth
This is the ability to dive into any module, regardless of its complexity or origin, to diagnose and fix a problem. It requires the humility to know when a situation exceeds one's capabilities and the willingness to learn whatever is necessary to resolve a critical failure. While fundamentals and breadth are gathered over time, just-in-time depth is a practice that must be honed through experience.
Navigating the Skeptic-Optimist Divide
The industry is currently split between two camps regarding agentic workflows. Optimists believe that agents will handle the heavy lifting, allowing humans to focus on higher-level architecture. Skeptics warn that relying on agents creates "haphazard mansions" of code that no human truly understands, leading to catastrophic failures.
Community discussion suggests that the reality lies between these extremes. While some argue that proof systems could provide mathematical guarantees for agent-generated code, others emphasize that the human role is shifting toward managing system boundaries—using sequence diagrams and architectural models to track inputs and outputs rather than individual lines of code. As one observer noted, the skill is knowing when to "go in a box" to debug a malformed message or logic error.
The Future of Engineering Expertise
Maintaining expertise in an era of automation is a significant, unresolved challenge. As engineers move further away from the metal, there is a risk that their ability to intervene effectively will atrophy. The most effective engineers of the future will be those who can maintain their "map" of the system while remaining capable of performing the deep-dive, hands-on work required when the agents fail. Ultimately, ownership is not about having written the code, but about being the person who can reliably navigate the system when the monsters emerge.