Research map / Multi-agent systems and games / LLM agents
STRATUS: A Multi-agent System for Autonomous Reliability Engineering of Modern Clouds
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes STRATUS, a state-machine-orchestrated multi-agent LLM system for autonomous cloud SRE that guarantees safety via Transactional No-Regression (undo-and-retry), outperforming SRE agents on AIOpsLab and ITBench mitigation tasks.
Problem
Autonomous mitigation of cloud system failures by LLM agents is risky because agent actions can worsen already degraded systems; existing SRE agents mostly assist humans and offer no safety or recovery guarantees.
Approach
A multi-agent system (detection, diagnosis, mitigation, undo agents) coordinated by a deterministic state machine, with a formal Transactional No-Regression safety property enforced via writer locks, sandboxed confinement, bounded transactions, and stack-based undo so failed mitigations are rolled back and retried from the initial state.
Key results
- STRATUS (GPT-4o) solves 69.2% (9/13) of AIOpsLab mitigation problems vs. 46.2% for the best baseline (AOL-agent GPT-4o), a 1.5x improvement.
- On ITBench mitigation, STRATUS (GPT-4o) solves 50.0% (9/18) vs. 9.2% for ITB-agent (GPT-4o), a 5.4x improvement.
- Ablation on AIOpsLab: removing retry drops success to 15.4%, and naive retry without undo reaches only 23.1%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use STRATUS (GPT-4o) instead of AOL-agent (GPT-4o) | AIOpsLab mitigation · AIOpsLab mitigation | success rate: 46.2% -> 69.2% | Helped |
| Use STRATUS (GPT-4o) instead of ITB-agent (GPT-4o) | ITBench mitigation · ITBench mitigation | success rate: 9.2% -> 50.0% | Helped |
| Enable retrying a different mitigation plan (ablation) | STRATUS (GPT-4o) · AIOpsLab mitigation | success rate: 15.4% -> 69.2% | Helped |
| Add undo and rollback to the original state when retrying (ablation) | STRATUS (GPT-4o) · AIOpsLab mitigation | success rate: 23.1% -> 69.2% | Helped |
| Enable retrying a different mitigation plan (ablation) | STRATUS (GPT-4o) · ITBench mitigation | success rate: 11.1% -> 50% | Helped |
| Add undo and rollback to the original state when retrying (ablation) | STRATUS (GPT-4o) · ITBench mitigation | success rate: 50% -> 50% | No clear effect |
| Set temperature to 0.2 instead of 0.0 (ablation) | STRATUS (GPT-4o) · AIOpsLab mitigation | success rate: 69.2% -> 51.3% | Hurt |
| Set temperature to 0.4 instead of 0.0 (ablation) | STRATUS (GPT-4o) · AIOpsLab mitigation | success rate: 69.2% -> 46.2% | Hurt |
| Set temperature to 0.6 instead of 0.0 (ablation) | STRATUS (GPT-4o) · AIOpsLab mitigation | success rate: 69.2% -> 56.4% | Hurt |
| Set temperature to 0.8 instead of 0.0 (ablation) | STRATUS (GPT-4o) · AIOpsLab mitigation | success rate: 69.2% -> 33.3% | Hurt |
Surprising findings
- On ITBench, disabling the undo agent leaves success rate unchanged at 50%, as many solved cases were resolved by pod restarts that exploit non-persistent injected faults.
- Naive retry without undo performs worse than undo-based retry, since failed mitigations further degrade the system state.
Limitations
- Perfect undo is hard to realize for application-specific state and external interactions; undo assumes exact state restoration.
- Rule-based confinement may not capture all destructive actions.
- Evaluation on small benchmark subsets (13 and 18 mitigation problems) with a single temperature-0 run per setting; results may be noisy.
Experimental setup
Datasets: AIOpsLab ITBench
Benchmarks: AIOpsLab ITBench
Metrics: success rate time steps monetary cost
Baselines: AOL-agent ITB-agent ReAct Flash
Builds on: CrewAI AIOpsLab ITBench ReAct
Closest papers by meaning
- Which Agent Causes Task Failures and When? On Automated Failure Attribution of LLM Multi-Agent Systems (ICML 2025 · similarity 0.83)
- Towards a Science of AI Agent Reliability (ICML 2026 · similarity 0.82)
- ITBench: Evaluating AI Agents across Diverse Real-World IT Automation Tasks (ICML 2025 · similarity 0.82)
- Rethinking the Reliability of Multi-agent System: A Perspective from Byzantine Fault Tolerance (AAAI 2026 · similarity 0.82)
- Why Do Multi-Agent LLM Systems Fail? (NeurIPS 2025 · similarity 0.82)
- ComplexMCP: Evaluation of LLM Agents in Dynamic, Interdependent, and Large-Scale Tool Sandbox (ICML 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.