AI usage patterns in software teams
AI-assisted coding is reshaping how software teams execute work, with more pull requests, automated changes, and agent-driven workflows—but many engineers report that gains in speed come with higher review, debugging, and maintainability costs. Commenters question whether activity metrics from tools like Linear meaningfully reflect customer value or ROI, raise concerns about data usage and privacy, and debate whether current LLMs can produce truly “great” code versus fast, average-quality output that shifts, rather than eliminates, human effort.
Developer experiences with AI coding tools
- Several posters say their workflow has become “LLM writes, human reads/cleans,” with long debugging and cleanup phases.
- Local LLMs are widely viewed as underwhelming compared to frontier/proprietary models.
- Some claim advanced agents with access to full codebases, type systems, and tests can rival or beat many humans at avoiding runtime errors, though not necessarily at avoiding bugs or poor designs.
- Others share success stories where autonomous agents iteratively instrument, test, and fix complex engine code over hours, producing visually correct results.
Code quality, “intent,” and limits of LLMs
- One camp argues LLMs fundamentally optimize for statistical patterns, so they struggle to produce “great,” scalable, maintainable code without strong human intent and constraints.
- Others counter that “intent” is just a set of optimization goals; in theory, with tight specs and enough compute, constraints could be satisfied without human-style intent.
- There is disagreement over whether LLMs really converge on the “statistical mean” or approximate a more complex distribution.
- Some argue that if users can articulate precise intent and checks, agents can produce high-quality code; “vibe-coding” prompts lead to trash.
Impact on workflows, metrics, and ROI
- Commenters note AI is used heavily in “how to build” (coding, research, architectural guidance bots) more than “what to build,” but this might not be visible in Linear’s metrics.
- AI-driven automations can bypass traditional ticket workflows, potentially reducing the importance of tools like Linear’s UI.
- More pull requests are seen as plausible due to easier small fixes and cleanups; but increased PR counts don’t equal increased value.
- Several criticize the article’s metrics as measuring what’s easy (PRs, usage) rather than outcomes like customer satisfaction, which tools like Linear can’t observe.
- There’s debate over ROI: many agree AI helps, but question whether the benefits justify large infrastructure bets and whether value goes to tool users or model providers.
Data collection and privacy concerns
- Some find it inappropriate that product usage data is mined and publicized, arguing it belongs to customers and may leak process insights.
- Others consider aggregated, anonymized stats harmless or even preferable to silent data sales, though the business value asymmetry is noted.