Agentic AI Between Capability and Reliability: A Structured Review of Architectures, Design Patterns, Evaluation Practice, and Enterprise Deployment

Autonomous software agents built on large language models have moved from research demonstrations to production systems in under three years. This review organises the evidence behind that shift and argues that the field’s binding constraint has changed. Capability is no longer the scarce resource; reliability under composition and governance under autonomy are. Drawing on peer-reviewed papers, preprints, standards documents, vendor engineering material, and analyst forecasts published between 2022 and mid-2026, we assemble a reference architecture for agentic systems, catalogue the design patterns that have stabilised around it, and trace how memory, planning, orchestration, and tool-invocation subsystems interact. We then confront a discrepancy that runs through the empirical record. On curated benchmarks, agents now approach or exceed human baselines: reported scores on SWE-bench Verified rose from roughly one fifth to four fifths of tasks solved in under a year, and operating-system interaction scores climbed from 12 per cent to above 72 per cent. On live interfaces and extended horizons the same systems complete between 7 and 50 per cent of tasks, and accuracy on long-horizon analysis falls by close to 47 points between early and late turns. We show that a single-parameter compounding model reproduces this gap without appealing to any additional failure mechanism, which reframes long-horizon fragility as an expected consequence of sequential dependence rather than an anomaly. Three further findings follow. Interoperability has converged on two complementary protocols, one for tool access and one for inter-agent messaging, yet neither encodes governance primitives such as authorisation, dissent, or deliberation. Commercial expectations are large and internally inconsistent, with forecasts of roughly one trillion United States dollars in annual agentic software spending by 2030 sitting alongside survey evidence that under a tenth of enterprises operate agents at scale and that two fifths of programmes are expected to be cancelled. Occupational effects to date resemble redefinition rather than substitution, with only one organisation in five reporting reduced service-desk headcount. We propose a Capability-Reliability-Governance framework for reading the literature, a minimum reporting standard for agent evaluation, and six falsifiable propositions for the research agenda through 2030.