Software Engineering: Moving From Coding to Leadership

Google Cloud's 2025 DORA report reveals that 90% of nearly five thousand technology professionals now integrate artificial intelligence into their daily workflows. A few years ago, a typical day for a software engineer involved opening files and manually typing code. Today, the profession is increasingly defined by describing tasks in natural language, auditing machine-generated output, and executing tests. As keyboard-centric tasks decrease, the burden of decision-making and accountability grows.

This shift makes the analogy of leadership particularly apt. Much like a project manager, the modern engineer is tasked with defining objectives, evaluating outcomes, and owning the final result. Coding remains, but its visible face has transformed.

What do the numbers indicate about AI in software development?

The Google Cloud DORA team dedicated their 2025 report to AI-assisted software development. Among the nearly five thousand technology professionals surveyed, 90% report using AI at work, with over 80% believing their productivity has increased. However, the same research indicates that approximately 30% have little to no trust in the resulting code. When viewed together, these figures clarify the current landscape: while the tools are ubiquitous and the sense of velocity is widespread, trust is not an automated feature. DORA characterizes AI as a multiplier: it strengthens high-performing teams while magnifying flaws in poor processes. The primary question is no longer whether tools exist, but whether a team’s habits regarding measurement, peer review, and platform hygiene are sufficient.

A similar narrative exists at the corporate level. Executives at Alphabet/Google noted in earnings calls that a significant portion of code is now generated by AI agents and subsequently vetted by human engineers. While these proportions have grown over time, the final human approval remains a mandatory step. In effect, the factory line of code generation has accelerated, but the quality control gate remains firmly on the human desk.

How does working with AI resemble leadership?

In a 2026 article for ACM Queue, Craig A. N. Soules addresses this transition with nuance. In routine applications, models often outperform humans in speed. In some environments, the human developer has ceased reading code line-by-line, opting instead for a model-to-model review process. Soules asks whether the developer is disappearing or if the invisible parts of the job are simply taking center stage. His answer favors the latter. Deciding what to build, assessing whether the outcome aligns with the intent, and accepting responsibility for failures are tasks that do not lend themselves to easy automation.

However, claiming that every developer has become a manager is incomplete. Soules identifies three critical breakpoints. First, a human subordinate has a reputation and a career, and can push back against incorrect requests; an agent has neither. Second, a high-functioning team will debate poor instructions, whereas current models tend to execute them with unearned confidence. Third, traditional management operates on a human tempo, typically measured in weeks. An agent can deviate in seconds, meaning that a "check it later" approach is insufficient. Monitoring must be continuous and measurable. Consequently, the new role is not a carbon copy of a manager. The scope is broader, the details are more granular, and the engineer must accept that the days of memorizing every line of code are effectively over. The work is beginning to resemble a control room, involving parallel streams, intermittent decision-making, and final validation.

What has fundamentally changed on the engineer's desk?

The practical change is that thinking no longer happens solely during the act of typing. Previously, ambiguity was often resolved through the process of writing code. Today, a vague prompt can faithfully produce a mountain of erroneous output. Clear specification is returning to the forefront, not in the sense of outdated, rigid documentation, but in terms of behavior: what the system must achieve, which constants must remain intact, and what must never occur. Deloitte’s framework on agentic software lifecycles supports this shift. The professional task has moved from writing lines of code to orchestrating multiple agents. Essential new skills include managing the context input, rapidly auditing agent outputs, and identifying the security risks introduced by agentic workflows. When providing instructions, simply saying "do this" is no longer sufficient; specifying what not to do, what to measure, and which files to avoid is now part of the process. When the output arrives, the question is not merely "does it work," but "did it solve the correct problem?" For those with low trust in AI, this skepticism is not laziness; it is a professional requirement.

What is overstated and what is here to stay?

The hype is well-documented. Not every team is working with a swarm of agents. In many environments, the workflow remains limited to chat interfaces and code editor suggestions. DORA notes that the agent-based mode of work is not yet a daily habit for everyone. What is permanent, however, is that as the cost of writing code drops, the value of review, measurement, and intent clarity rises. Increased velocity does not automatically equate to higher quality. Producing code with AI does not mean retiring the keyboard; it means treating the work as an act of delegation and ownership, where the constant need for oversight remains because the agent never objects.

Notes and Sources

1. Soules, C. A. N. (2026). "Where to Draw the Line". ACM Queue, 24(4). Regarding the automation of software engineering, behavioral specification, and continuous oversight. queue.acm.org.
2. DeBellis, D. et al. (2025). DORA 2025 State of AI-assisted Software Development Report. Google / DORA. Survey of nearly 5,000 respondents; 90% AI adoption; ~30% low/no trust in generated code. Summary: Google Cloud Blog; report record: research.google.
3. Alphabet/Google earnings calls and reports: regarding the high percentage of code generated by AI agents and reviewed by engineers (with reporting suggesting levels rose from the 30% range to approximately 50%). See Computer Weekly coverage: computerweekly.com.
4. Deloitte (2026). Future of Engineering: Agentic AI in software engineering — from SDLC to AO-DLC. Regarding agent orchestration, context engineering, and agent oversight. PDF.