Failures: Deterministic Failure-Mode Guardrails Make Coding Agents Build Resilient Systems

This paper investigates whether deterministic failure-mode analysis can improve the resilience of software produced by AI coding agents. We present Failures, a Model Context Protocol (MCP) server that encodes eleven software reliability dimensions, including atomicity, idempotency, timeout and ambiguous-outcome handling, concurrency, ordering, consistency, availability, resource exhaustion, recovery, observability, and retry safety. Failures provides deterministic checks with line-level evidence and guides coding agents to reason about failure modes before implementation and verify completed code afterward. We evaluate the approach across six realistic software scenarios using controlled comparisons and blind agent studies involving Claude, Cursor, and Codex. Across 17 matched agent comparisons, 16 improved on at least one resilience criterion. We additionally evaluate six adversarial cases to identify limitations and false positives in the deterministic analyzer. The work examines how AI-assisted software development can move beyond functional correctness toward software that remains reliable under failures, ambiguity, retries, and concurrency. FILES- Failures_MCP_Paper.pdf — the paper (15 pp., 2 figures, 10 tables, 16 references).- failures-artifact.zip — evaluation/ + examples/ + mcp_server/, the complete artifact needed to reproduce Tables 7, 8 and 9 of the paper. REPRODUCING THE RESULTS pip install -e . (or: pipx install failures-mcp) python evaluation/run_evaluation.py --mode proxy python evaluation/run_evaluation.py --mode manual Both commands run offline with no external services. The proxy run reproducesthe 6-scenario study (Table 7); the manual run scores the frozen agent outputsin evaluation/runs/{claude,cursor,codex}/ and reproduces Table 8 (17 matchedbaseline/failures-enabled pairs; 16 improve on at least one resiliencecriterion). evaluation/results.json and evaluation/adversarial_results.json arethe machine-readable sources for every table. Scenario prompts are hashed andthe hash plus repository commit are recorded in each run manifest. CODE https://github.com/mayowa-kalejaiye/Failures CITATION Kalejaiye, Oluwamayowa (2026). Failures: Deterministic Failure-Mode Guardrails Make Coding Agents Build Resilient Systems. Zenodo. https://doi.org/10.5281/zenodo.22966362 LICENSE CC BY 4.0. The software artifact is also released under the MIT License.