At a Glance
Most research on cloud-native, edge-native and agent-driven AI systems is still stuck at prototypes or simulations; only a small share shows production-proven reliability, leaving real-world edge and agent deployments under-supported.
ON THIS PAGE
What They Found
A large mapping of 1,794 studies from 2017–2026 shows most work is done in labs: two-thirds of determinable studies used prototypes or testbeds, fewer ran real production data or deployments. Where production evidence exists, it shows successful techniques for preventing cascading failures, doing trace-based root-cause analysis, predictive autoscaling, and reducing cold-start problems. Edge-native systems and physical AI mostly rely on simulations and testbeds, and multi-agent or agentic systems have some safety and repair techniques but remain fragile in real deployments. Major gaps are production evidence for edge systems, shared resilience benchmarks, and dependable operations for agent-driven apps.
Not sure where to start?Get personalized recommendations
By the Numbers
11,794 primary studies (2017–2026) and 95 secondary studies were analyzed in the mapping review.
2Of 1,003 studies with a clear deployment level: 66.2% used prototypes or testbeds, 18.0% used simulations, and 15.8% used production data or live deployments.
3A focused synthesis of 93 high-relevance studies found production-proven solutions for cascading-failure prevention, trace/graph root-cause analysis, predictive autoscaling, and cold-start mitigation—but few production examples for edge-native or agentic systems.
What This Means
Platform engineers, SREs, and technical leaders deciding whether to run AI workloads at the edge or deploy agentic systems should care: the literature shows what is proven in production and where risk remains. Researchers and tool builders should use these gaps to prioritize shared benchmarks, production monitoring, and techniques that make multi-agent systems dependable.
Ready to evaluate your AI agents?
Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.
Learn MoreKeep in Mind
The synthesis core was appraised from indexed abstracts, so some production evidence may be under-reported if not captured in abstracts. Findings reflect only the documented corpus through the review date; emergent deployments after that are not included. Distribution of study types varies by topic, so some subareas (edge, physical AI, agentic ops) have especially thin production evidence.
Methodology & More
The study ran a systematic mapping review following established review steps, searching literature across cloud-native systems, edge AI, and agentic AI topics and adding an earlier set of records. After deduplication and automated filtering, 1,794 primary and 95 secondary studies from 2017–2026 were classified into a 13-category taxonomy and graded by evidence level (simulation, prototype/testbed, production data, production deployment, or unclear). A 93-study synthesis core was manually appraised from abstracts and dataset decisions are released for transparency. Results show a field still dominated by prototypes and testbed experiments: of 1,003 studies with a determinable deployment level, 66.2% were prototypes/testbeds, 18.0% simulations and only 15.8% used production data or live deployments. Concrete production-proven techniques do exist—examples include systems that prevent cascading failures, graph- and trace-driven root-cause analysis, predictive autoscaling, and production trace evaluations of cold-start mitigation. By contrast, edge-native and physical AI research remains largely simulation-driven, and agentic systems have promising tool-based diagnostics, explicit safety rules, trace-driven repair, and workload-aware scheduling but are empirically fragile. The paper highlights three priorities for the community: more production evaluations (especially for edge), shared resilience benchmarks and evaluation patterns, and better practices for continuous monitoring and tracking of agent behavior and failures.
Avoid common pitfallsLearn what failures to watch for
Credibility Assessment:
Single author with low h-index (2) and arXiv venue; limited signals of established reputation -> emerging/limited info.