The AI Productivity Gap
AI coding tools and agents are changing how software engineers spend their time, compressing implementation work but often expanding planning, supervision, and review. Many report gains in individual throughput—especially for juniors and well-structured tasks—but also complain of “babysitting” agents, bloated or subtly broken code, and degraded understanding of systems. A recurring theme is that without deliberate process and quality controls, AI shifts bottlenecks rather than removing them, raising questions about long‑term team productivity, skills development, and codebase health.
Perceived Productivity Gains
- Experiences range from large speedups (2x to “man‑week in 1.5 hours”) to neutral or negative overall productivity.
- Many note AI drastically compresses implementation/boilerplate time but does less for architecture, integration, and validation.
- Some leaders mistakenly treat AI as a uniform X% speedup; commenters argue improvements are highly uneven across tasks.
Shift in Developer Work
- Several describe becoming coordinators or “herders of cats” for multiple agents instead of hands‑on coders.
- Babysitting, re‑prompting, and steering agents can feel tedious and mentally draining.
- Some limit AI use to autocomplete or initial boilerplate to avoid this overhead.
Code Quality, Review, and Bugs
- Strong concern that AI code is less trustworthy, longer, more abstracted, and harder to reason about.
- Bugs tend to be “non‑normal”: odd deletions, monkey patches, strange concurrency issues, subtle cheats.
- Review time often increases due to bloated PRs, verbose test notes, and authors who understand their own code less.
- Others argue human review is even more essential with AI; a minority suggests human‑centric style expectations may be mismatched to AI‑generated code.
Process & Tooling Matters
- Some teams report ~2x productivity with strict process: upfront specification, planning, adversarial review by a second model, and heavy documentation.
- Emphasis on “harness engineering,” context management, standardized onboarding/wrap‑up, and knowing when to kill or redirect bad agent sessions.
- Others see chaos when AI is dropped into old workflows without process redesign.
Juniors, Seniors, and Skills
- Juniors and interns can produce much more output with AI, but there is concern they develop shallow understanding and “LLM overseer” habits.
- Some fear long‑term erosion of expertise and reliance on a few vendors.
- Suggested workflows: beginners manually retype AI proposals to learn; experienced devs use AI mainly for completion, not architecture.
Organizational & Systemic Limits
- Commenters invoke Amdahl’s law / theory of constraints: speeding coding alone mostly grows the work queue.
- Upstream (requirements) and downstream (QA, ops, adoption) processes remain serial and human‑bound.
- Small teams and carefully designed “AI‑friendly” stacks may see outsized benefit; at larger orgs, gains are harder to realize and measure.