Updated 17 October 2026 · Testing and evaluation
Abstract. Generative AI increases code-production capacity faster than it increases confidence. Quality engineering must therefore shift from test execution to evidence orchestration: deterministic checks, model-based evaluators, security probes, human judgment and production telemetry are combined in bounded repair loops. The goal is not self-healing software at any cost; it is verifiable recovery without silent regression.
The oracle problem becomes central
An LLM can generate tests that compile and still encode the same misconception as the generated implementation. Empirical evaluation across Java projects found that prompt configuration materially affects unit-test generation and that capability varies across open and commercial models (Yang et al., 2024). Generated tests therefore need independent oracles: specifications, properties, differential behavior, metamorphic relations, security invariants and observed production contracts.
| Evidence layer | Examples | What it can establish |
|---|---|---|
| Static | Types, lint, dependency and policy scans | Structural conformance |
| Deterministic dynamic | Unit, integration, property and contract tests | Observed behavior under specified cases |
| Adversarial | Fuzzing, mutation, red-team prompts, fault injection | Robustness against variation and abuse |
| Model evaluator | Rubric grading, semantic comparison, critique | Probabilistic quality signals |
| Human | UX, architecture, legal and ethical review | Contextual judgment and accountability |
| Production | SLOs, traces, incidents and user outcomes | Real-world fitness |
Evaluator engineering
An evaluator is a measurement instrument and must itself be evaluated. Track agreement with expert labels, sensitivity to prompt phrasing, false-accept cost, false-reject cost and drift across model versions. Use blinded pairwise comparison when absolute scoring is unstable. Keep critical gates deterministic wherever possible and reserve model judges for dimensions that genuinely require semantic interpretation.
RAGAS illustrates decomposed evaluation for retrieval systems rather than a single “answer quality” score (Es et al., 2023). The same principle applies to agents: separately score plan validity, tool selection, action safety, evidence completeness and final outcome.
The self-healing loop—with brakes
- Detect a failing invariant and capture a minimal reproducible state.
- Localize likely causes without modifying production.
- Generate multiple repair hypotheses and predicted side effects.
- Apply one candidate in an isolated environment.
- Run focused tests, then the retained regression obligation set.
- Evaluate security, performance and architecture impact.
- Promote only with required approval and a tested rollback.
LoopsBench retains completed graph nodes as regression obligations, making regressions visible during long-horizon execution (Li et al., 2026). SWE-Gym similarly treats executable environments and verifiers as training infrastructure for software agents (Pan et al., 2024).
Testing the agent, not only its code
Agent test suites should include prompt injection, poisoned tool output, ambiguous requirements, partial permissions, expired credentials, network failure, cost exhaustion, cancellation and human override. Assert that the agent stops safely. Measure unauthorized-attempt rate, escalation precision, recovery success, trace completeness and cost per verified task.
SWE-bench remains influential because it evaluates real repository issues rather than toy completions (Jimenez et al., 2024). Yet benchmark optimization can overfit public tasks; maintain private, time-sliced and organization-specific suites. Compare end-to-end outcomes and reviewer effort, not vendor leaderboard numbers alone.
Agentic QA market architecture
Codex, Claude Code and other coding agents can write tests, execute suites and review diffs. Dedicated observability and evaluation platforms add trace analysis, datasets and experiment comparison. The strategic design is a provider-neutral evidence ledger: every accepted change links to tests, model/version, source context, approvals and deployment outcome. This allows an organization to change tools without losing quality history.
A minimum quality gate for AI-authored change
- Acceptance criteria trace to executable tests.
- New behavior has positive, negative and boundary cases.
- Mutation or property testing challenges test adequacy where risk justifies it.
- Security and dependency checks run in the same immutable build.
- A reviewer can distinguish agent-authored assumptions from verified facts.
- Rollback is defined and exercised for consequential changes.
Quality principle. A self-healing loop is trustworthy only when it can prove what it preserved, expose what it could not verify and terminate safely when evidence is insufficient.
References
- Es, S., James, J., Espinosa-Anke, L., & Schockaert, S. (2023). RAGAS: Automated evaluation of retrieval augmented generation. arXiv. https://arxiv.org/abs/2309.15217
- Jimenez, C. E., Yang, J., Wettig, A., Yao, S., Pei, K., Press, O., & Narasimhan, K. R. (2024). SWE-bench: Can language models resolve real-world GitHub issues? International Conference on Learning Representations. https://arxiv.org/abs/2310.06770
- Li, H., Fang, Z., Feng, R., Zhao, Y., Liu, J., Gao, P., Ye, H., Lin, D., Lin, Q., Rajmohan, S., & Zhang, D. (2026). LoopsBench: From harness engineering to loop engineering in coding agent evaluation. arXiv. https://arxiv.org/abs/2608.00267
- Pan, J., Wang, X., Neubig, G., Jaitly, N., Ji, H., Suhr, A., & Zhang, Y. (2024). Training software engineering agents and verifiers with SWE-Gym. arXiv. https://arxiv.org/abs/2412.21139
- Yang, L., Yang, C., Gao, S., Wang, W., Wang, B., Zhu, Q., Chu, X., Zhou, J., Liang, G., Wang, Q., & Chen, J. (2024). On the evaluation of large language models in unit test generation. arXiv. https://arxiv.org/abs/2406.18181