AI coding agents generate more code, but not more software
AI coding agents are dramatically increasing the amount of code and tests engineers can generate, but many argue this isn’t translating into more high‑quality, shipped software at companies. Commenters point to human bottlenecks like code review, unclear requirements, organizational process, and the difficulty of validating AI-written changes at scale—especially in large, regulated or legacy environments. At the same time, several note a real boom in personal and small‑scope tools built with AI, suggesting the main constraint is less the technology and more how organizations choose to structure and govern software work.
Effect on productivity and “more software”
- Many report writing far more code and small tools, especially for personal projects and homelabs. Things that felt “not worth the effort” are now feasible.
- In enterprises, perceived net throughput gains are modest: coding is a small fraction of engineering time; design, analysis, and coordination dominate.
- Some argue the linked study’s data (mostly 2025) understates newer models; others say capabilities haven’t meaningfully changed the bottlenecks.
Code review and quality bottleneck
- Strong consensus that review, not typing speed, is now a main bottleneck.
- Opinions split:
- Some find traditional line‑by‑line review tedious and want heavy automation or to skip it for “non‑serious” changes.
- Others insist review is essential for shared understanding, maintainability, accountability, and catching design flaws that tests miss.
- Hybrid models are emerging: humans focus on architecture and intent; agents do first‑pass linting, correctness checks, and performance/security scans.
Tests, tooling, and maintenance
- Agents generate huge volumes of tests and comments, often verbose, repetitive, and white‑box.
- Concerns: “test slop” consuming CI resources, tests encoding existing bugs as “correct,” and difficulty triaging when hundreds fail at once.
- Some use agents to audit and refactor test suites, with mixed but sometimes impressive results.
Specifications, design, and intent
- Recurrent theme: AI is weak at validating intent and UX; strong at local correctness if requirements are explicit.
- Calls for better human‑written specs, BDD/user stories, and spec‑driven development and review. Others warn big upfront specs historically failed and iteration still matters.
- Worry about “AI all the way down”: AI‑generated requirements → specs → tests → code → reviews, with no human really understanding the system.
Personal vs enterprise usage
- Many anecdotes of solo/side projects, custom tools, and even complex agent harnesses that “would never have existed” pre‑agents.
- Enterprise teams report slower gains due to legacy processes (Jira, long backlogs, risk/compliance, politics) and leadership hype demanding unrealistic velocity.
Cultural and organizational factors
- Broad agreement that coding was never the main bottleneck; communication, product vision, and process are.
- AI acts as a force multiplier: disciplined teams get faster and better; “cowboy” or process‑theatrical teams just produce more low‑quality slop.
- Open questions: how to replace/reshape code review, how far to trust autonomous agents, and whether current practices can scale with “infinite cheap code.”