Ask HN: If AI tools are so good, why isn't your bug backlog empty?
AI coding tools are dramatically accelerating software output, but engineers report mixed results on whether this actually clears bug backlogs. Some teams and solo developers say their bug queues are effectively empty and credit AI for faster fixes and refactors, while others find models unreliable, prone to shallow or breaking fixes, and hard to trust on large or critical systems. Many note that bottlenecks have simply shifted—from typing code to specifying requirements, reviewing changes, validating behavior, and deciding which bugs or features are worth addressing at all.
Overall split in experiences
- Many report bug backlogs that are empty or nearly so; others say their queues are as full as ever.
- Where backlogs are light, AI is framed as a strong multiplier, but not the sole reason (good testing, small focused systems, disciplined engineering also matter).
- Several say bug queues shrank, but feature/idea backlogs grew dramatically.
Cases where bug backlogs are (nearly) empty
- Solo developers and small teams using strong models + tests say:
- Most clearly specified bugs are auto-investigated, fixed, and regression-tested by agents.
- Output and refactors they’d postponed for years were cleared in weeks.
- Remaining items tend to be “nice-to-haves,” performance tweaks, or ambiguous edge cases.
- Some organizations in stable domains (financial, line-of-business apps, devops infra) report:
- No urgent tech debt.
- Backlogs of essential infra work emptied.
- Constraints shifting to product definition, deployment, sales, and user adoption.
Why many backlogs are not empty
- AI often fixes symptoms locally instead of eliminating entire classes of bugs.
- Hard bugs remain: heisenbugs, rare C++ crashes, huge legacy codebases, underspecified issues.
- Some say AI frequently introduces regressions or bizarre implementations, eroding trust.
- Verification is costly: human review, regulatory constraints, extensive testing, and acceptance testing remain bottlenecks.
Limits of current AI tools
- Context and architecture: agents lack persistent, project-wide mental models; AGENTS.md and custom documentation help but aren’t enough for very large systems.
- Training incentives favor “defensively correct” code, not minimal, maintainable designs.
- Behavior is described as inconsistent: brilliant on some tasks, inexplicably bad on others.
- AI is better at finding issues than safely fixing them at scale.
Organizational and structural factors
- Many teams still prioritize new features over bug work; AI just amplifies that choice.
- Human processes—code review, governance, politics, unclear requirements—dominate timelines.
- Some argue a deep bug backlog signals broader disorganization AI cannot fix.
- Concepts like Parkinson’s Law and Jevons paradox are invoked: more efficiency yields more code, features, and ultimately more bugs to manage.