AI Coding: Harder Engineering

The rapid adoption of artificial intelligence in software development prompts critical questions beyond mere productivity gains. While AI promises faster coding, it concurrently elevates the complexity of software engineering tasks. This shift necessitates heightened discipline and extensive knowledge from developers, transforming the nature of their work and the expectations placed upon them.

Simon Willison, a respected voice in the developer community, articulates this evolving challenge clearly. He states that while AI coding agents accelerate development, they paradoxically make software engineering more difficult. Willison emphasizes that extracting the full potential from these tools demands “extraordinary discipline and knowledge.” This perspective underscores that AI does not reduce the need for human expertise; it redefines it.

Accelerated Development, Increased Responsibility

The idea that faster coding does not equate to easier work is gaining traction among developers. Hillel, a reader responding to Willison’s observations, concisely stated, “It doesn’t get easier, you just get faster.” This sentiment highlights a crucial distinction: while AI can dramatically increase implementation speed, it also expands the scope and complexity of projects that teams undertake. This expansion introduces new layers of decision-making, verification, and oversight.

Experienced operators, such as Geoffrey Huntley, describe working with AI coding tools as “incredibly taxing.” Even with significant productivity boosts, individuals find themselves dedicating extended hours to manage projects. This intensity often stems from an increased capacity to attempt more ambitious projects, rather than a reduction in effort. The enthusiasm developers show for what AI makes possible often leads to a willingness to take on more.

AI coding agents demand skilled software engineers. They do not eliminate the necessity for engineering discipline; instead, they amplify the consequences of its absence. Having the necessary expertise is one aspect, but maintaining sufficient capacity to apply that expertise effectively becomes another challenge. For instance, a team leveraging AI to expedite a postponed migration might encounter numerous decisions regarding compatibility, potential customer disruption, and the preservation of existing functionalities. While AI handles the coding, engineers dedicate more time to resolving these intricate questions.

The gains provided by AI are undeniable, yet they come with increased responsibilities. Comparing a new, more ambitious workload to a previous one and concluding that engineering has become simpler overlooks the true cost. The time saved through AI is often reinvested into addressing more complex problems. The effectiveness of AI agents in reducing correction and verification work can also enable projects that were previously impractical. The critical organizational choice lies in determining how much of this newfound capacity becomes breathing room and how much transforms into additional commitments.

Elevated Expectations and Workload Expansion

New research supports the concern that productivity gains can lead to unsustainable workloads. Aruna Ranganathan and Xingqi Maggie Ye, researchers at UC Berkeley, conducted an eight-month study at a 200-person technology company. Their qualitative research, involving observations and over 40 interviews across various functions, revealed a notable trend. Employees expanded their responsibilities, integrated AI prompting into previously idle moments, and managed more tasks concurrently. This expansion was largely voluntary, driven by excitement over new capabilities.

The primary issue identified was how this initial enthusiasm reset expectations. As Ye explained, “what was once extra effort becomes standard performance.” While this qualitative research from a single company does not represent a universal truth, it strongly indicates how a surge of experimentation can normalize into an ordinary workload against which all employees are measured. This phenomenon suggests that initial productivity boosts may not translate into sustained, manageable improvements if not properly managed.

Management faces the challenge of accurately defining success metrics in an AI-augmented environment. If the objective is to undertake more ambitious work, organizations must acknowledge that teams are allocating their gains toward this ambition. If cost reduction is the goal, then the full cost of project completion and maintenance must be meticulously measured before assuming savings. If engineers are working longer hours to meet new targets, a portion of the apparent improvement is simply a result of increased human labor.

The proliferation of tools designed to manage coding agents highlights this evolving landscape. While better tools can assist developers in recovering context and completing tasks, they do not dictate the amount of work management should expect. An efficiently organized queue can still contain an overwhelming volume of work. Organizations must ensure that the “gains” from AI are not merely disguising an escalating workload.

Strategic Allocation of Productivity Gains

Willison’s own experiences provide a practical illustration of how AI can be productively integrated into workflows. He detailed a security audit of Datasette in September, where multiple coding agents were instrumental. Prompted by external vulnerability reports, the audit involved successive rounds of problem discovery. Willison and Alex Garcia collaboratively divided the tasks: one wrote tests to demonstrate issues, while the other implemented fixes. Each vulnerability underwent examination by two human engineers, augmented by agents using different AI models. Corrective measures were integrated into the main development branch, with crucial changes also applied to the stable release.

In this scenario, AI facilitated the discovery of valuable work, and a clear, explicit process was established for its completion. The additional findings proved valuable because the maintainers took concrete action on them. This account emphasizes that merely quantifying the number of hours saved by AI overlooks a significant portion of the overall accomplishment. A longer list of vulnerabilities to address signifies a more thorough audit, making it illogical to label it a failure due to the additional work it generated. Similarly, assuming that the individuals responsible for implementing these fixes suddenly became less essential would be equally illogical.

For engineering leaders, this implies a need for more insightful discussions than simply demanding a higher percentage of code generated by AI. Leaders should inquire about how the saved time was utilized. Did a product enhancement reach customers more quickly? Did the efficiency gained in implementation enable a more comprehensive audit? Did review processes extend into evenings? These outcomes represent distinct results, even if the coding agent’s performance appeared consistently impressive across all scenarios.

Furthermore, engineers should be empowered to allocate some of the productivity gains toward making subsequent tasks less demanding. Developing a reliable test to prevent recurring failures, creating clearer interfaces, or eliminating unnecessary dependencies can reduce the number of decisions that require revisiting. Such foundational work demands inclusion in project plans. Without this deliberate allocation, every enhancement in implementation speed risks being consumed by the relentless pursuit of new features, while the accumulating cost of understanding and maintaining the system continues to grow.

An enthusiastic experimental phase should not automatically translate into staffing assumptions for the next quarter. Before institutionalizing a burst of output as a standing commitment, it is essential to determine whether the team could sustain that level of productivity within its normal workday, encompassing review and maintenance tasks. Individuals may choose to immerse themselves in engaging projects, but this does not establish a baseline for the amount of work they can routinely absorb. This also should not become an obligation for colleagues who did not volunteer for the initial experiment.

Willison’s perspective describes demanding work that is ultimately worthwhile. This presents a far more credible argument for AI than promises of effortless software engineering. Enterprises should actively pursue the enhanced capabilities offered by AI but must also honestly budget for the human resources and time required to effectively utilize these tools. The productivity gain should be integrated into the plan only once. If an expanded project roadmap is only achievable because engineers consistently extend their workdays, then a portion of that gain is simply derived from increased human labor, with AI acting merely as an accelerator rather than a true enabler of sustainable efficiency.