Gergely Orosz, author of 'Practical Engineer', outlines a series of ongoing phenomena in the software industry, IDE decline, and increased engineer workload
Gergely Orosz, author of 'Practical Engineer', outlines a series of ongoing phenomena in the software industry and asks peers what has been overlooked. His background includes engineering roles at Uber and Skype, with a focus on practices in large companies and startups. Key points align with recent communication themes: declining use of integrated development environments (IDEs), token consumption no longer seen as a badge of honor, code reviews overwhelmed by generated output, open-source models used for cost savings and data locality, shrinking middle management, increased workload, and difficulty hiring outside top labs.
Orosz has personally documented the decline of IDEs: after about twenty years primarily working in editors, he has submitted a large volume of work in the past two weeks without opening VS Code or IntelliJ, instead using command line, web, and mobile to prompt code generation. A JetBrains survey from May to July 2026 indicates that about 90% of professional developers use coding agents at least weekly, with 68% using them daily; Claude Code penetration is about 39%, approximately 47% in the U.S., while GitHub Copilot dropped from about 29% to 21%, and Cursor's adoption fell from about 18% to 12%. The focus of tools has shifted from plugin completion to agent sessions.
The exit of token leaderboards corresponds to billing discipline. A pulse on September 3 stated that migrating simpler workloads to open-source models is the easiest way to save about half of AI expenses while avoiding data transfer. A September 8 article questioned whether the volume of generated output by 2026 would exceed human tracking capabilities, and whether the decades-old differential review process is being transformed or is merely nominal. An internal document from Anthropic noted that testing and review are increasingly delegated to models, while the "two-pizza team" concept remains.
Management and human resources are read together. An August 18 communication noted an increase in departures among CTOs, engineering VPs, and engineering leads, with high-status positions being voluntarily relinquished. Hiring remains difficult, attributed to non-top labs: agents increase per capita output but do not automatically expand the pool of qualified candidates. The increased workload stems from the same mechanism—agents raise the upper limit of tasks that can be accepted, filling schedules rather than reducing weekly working hours.
The market mechanism is a redistribution of engineering staffing and tool billing. Buyers are companies that need to maintain delivery with fewer middle managers; sellers are agent vendors and open-source model hosts. Funding is shifting from seat-based IDE licensing to token-based and self-built agent inference, and from middle management salary packages to a few senior engineers who can command agents. Beneficiaries are platform teams that can cut their bills in half using open-source models; those under pressure are mid-level positions that survive on line counts and the old IDE business model that charges for completion plugins. The driving events are observation posts and communication topics, not an official employment report.
The volume of generated output first overwhelms reviews, then forces tools away from editors. The saved tokens do not translate into rest but rather into more tasks scheduled in the calendar.
Source: Public information
ABAB AI Insight
Orosz writes the on-site feelings into a verifiable communication directory. The decline of IDEs is not the disappearance of editors, but a shift in submission entry to agent command lines; JetBrains' penetration rate supports this feeling with a survey. The exit of the token leaderboard indicates that CFOs have reviewed the bills, with open-source models being viewed as the cheapest half of inference. The nominal existence of reviews is due to differential views being designed for human time, while agents generate patches by the minute. The simultaneous reduction of middle management and increased workload occurs because the coordination layer is dismantled, and task funnels directly feed into those who can write prompts.
The capital path is a relocation of tool budgets. Copilot-style seat fees give way to Claude Code and self-developed agents; companies like Ramp create Inspect to reclaim vendor gross margins internally. Open-source models keep data in proprietary clouds, with compliance and costs on the same invoice. The wave of departures among engineering leads indicates that equity and status are being exchanged for rest, showing that the first costs of the agent revolution fall on those who organize others, not on model labs. Hiring difficulties arise because companies seek individuals who can design agent workflows, while supply continues to recruit based on traditional job levels.
Analogies can be drawn to the transition from assembly to high-level languages and cloud computing compressing data center staffing into platform teams. This transition is faster: command line agents have surpassed plugins within months. The industry phase is about controlling orchestration rights. Those who can connect repository events to agent clusters can deliver more differences with fewer middle managers. Top labs feed their engineers with internal models, while other companies save half the money using open-source models, widening rather than closing the talent price gap.
Structural changes signify technological substitution. The mechanism is that once generation speed exceeds human reading speed, old quality control positions lose their measurement units, and new quality control becomes sampling and testing agents. Increased per capita output raises task quotas; the increased workload is a function of quotas, not model failures. The scarcity of excellent engineers is due to the shift from writing statements to designing systems that can be agentized.
ABAB News · Cognitive Laws
- Once generation speed exceeds reading speed, reviews will first run in place.
- Saved tokens will turn into new tasks, not vacations.
- After middle management thins out, those still at the funnel's mouth bear the burden.