LLM Multi-Agent Systems for Self-Healing Distributed Systems: A Critical Review of Architectures, Operational Autonomy and Evaluation Evidence

Incidents in large cloud services are costly, and most of the expert effort they absorb goes into diagnosis; repair usually draws on a short list of routine actions. This critical review appraises large language model (LLM) agent systems for self-healing distributed systems, drawing 32 core sources (26 systems and six benchmarks) from a corpus of 208 and assessing each against fixed questions on the lifecycle stage evaluated, the evidence substrate, the safety mechanisms at the execution boundary and the comparability of reported results. When each system is placed on a six-stage capability ladder at the highest stage its authors actually evaluated, just over half stop at diagnosis. None of the six systems evaluated at autonomous execution was tested on production incidents; all used a benchmark, testbed or simulation, and few report post-action verification or rollback. Mapping agent failure modes against reported safeguards shows that repeated action loops and failed termination are each addressed by a single system. Reported results cannot be compared across studies: three evaluations on the same public benchmark used three metrics and three model backbones, and most gains are measured against each team’s own baselines. On this record, LLM agents are supported as diagnostic assistants behind approval gates. Comparable evidence on autonomous remediation will need a shared protocol that fixes and reports the backbone, scores each lifecycle stage, replays matched incidents and discloses cost.