Software Quality with Agents
Agent-assisted development does not eliminate verification; it changes its placement and volume. Across the corpus, robust practice combines deterministic checks, independent review, explicit triage, and evidence that a person can inspect—not speed alone.
Synthesis
- TanStack’s provider matrix uses deterministic fixtures, browser tests, feature flags, and isolated test routes to make cross-provider behavior reproducible on each pull request. [src]
- Harness engineering separates fast computational controls such as tests and linters from slower inferential review; Yegge and Lawson both argue for repeated review or multiple reviewers rather than treating first-pass output as complete. [src] [src] [src]
- Coinbase’s hiring account adds a human capability dimension: evaluate effective AI use, understanding of limits, repository-level debugging, and judgment—not merely whether an applicant can solicit generated code. [src]
Coverage expansion
David Gasquez argues that unreliable agent components can still participate in reliable systems when the surrounding design uses narrow contexts, repeated and diverse judgments, explicit aggregation, visible failures, and recovery paths. His pairwise-comparison example is an engineering proposal and experiment narrative, not evidence that every ensemble design is superior. [src]
This system-level reliability framing connects to Agent Harness Design and the incident evidence practices in Incident Debugging And Observability.