Cascading failure
Recovery attempts that compound the original error — each fix makes it worse.
a.k.a. recovery-spirala.k.a. compound-errora.k.a. error-cascadea.k.a. fix-regression
Description
The agent detects an error and attempts a fix that introduces NEW errors; repeated iterations spiral further from the working state. A simple failure becomes an irrecoverable tangle after N recovery attempts. Distinct from logic_error (which is about the initial bug itself) and from context_pollution (which is about session-level state drift rather than error-recovery compounding).
Detection signals
- Number of files modified grows per recovery iteration.
- Error messages change category across attempts (e.g., ImportError -> TypeError -> AttributeError).
- Same file is touched N times in rapid succession with diverging diffs.
- Test pass count decreases monotonically after iteration 2.
Prior work
- Mehil B. Shah (2026) — Characterizing Faults in Agentic AI: A Taxonomy of Types, Symptoms, and Root Causes
- Melis Cemri (2025) — Why Do Multi-Agent LLM Systems Fail? MAST: A Multi-Agent System Failure Taxonomy
Crosswalks
Agent's fix attempts introduce new errors that compound.
DAPLab "Cascading Error Recovery" is the direct counterpart.
Agent cycles over the same state or tool call indefinitely.
MAST "Infinite Loop" is one manifestation of a recovery-spiral pattern.
Related failure modes
- Logic error — Correct API use, incorrect logic.
- Context pollution — Trajectory drift, hallucination spirals, overlong or off-task context.
Mitigating tools
- bifrost — Fastest enterprise AI gateway with adaptive load balancing, failover, and guardrails.
- coze-loop — Full-lifecycle AI agent management platform with debugging, evaluation, and monitoring.
- lmnr — Laminar — open-source observability platform purpose-built for AI agents.
- manifest — Smart model routing for personal AI agents.
- sentrux — Real-time architectural sensor that scores codebase modularity, acyclicity, and redundancy as the agent writes.