Agentic AI: the autonomous tester we dreamed of?

Imagine software that doesn't just follow orders but decides on its own how to hunt down the most hidden bugs. That's the promise of Agentic AI, a new generation of artificial intelligence capable of autonomy and initiative. A systematic study sifted through 37 research papers to see where we stand.

The good news? These smart agents are already acing "Product Assurance": they generate tests autonomously, hunt defects, and manage reports with surgical precision. They often work in teams (multi-agent architectures) and show moderate, goal-oriented autonomy. Result: skyrocketing efficiency, fewer human errors, and more maintainable code.

But it's not all smooth sailing. The study highlights three big "buts": limited scalability, difficulty generalizing across different projects, and, above all, low transparency. We often don't understand *why* an agent made a specific decision: a black box that's scary in critical contexts.

There's another gap: these digital helpers support coders and testers (execution level), but almost completely ignore project managers and CTOs. A holistic process view is missing.

The path to fully autonomous, trustworthy QA is mapped out, but we need research on validation, explainability, and decision support at the top. Only then will Agentic AI become the colleague we trust blindly.

The study, conducted following PRISMA guidelines and Systematic Mapping methodology, analyzed 37 primary studies extracted from five major databases (IEEE Xplore, ACM Digital Library, ScienceDirect, Scopus, Web of Science). Classification used a framework based on three key Agentic AI dimensions: behavior (goal-oriented, diagnostic, collaborative), architecture (single-agent, multi-agent, hybrid), and autonomy level (low, medium, high).

Results show a strong concentration (over 60%) in Product Assurance activities, specifically automatic test case generation (model-based, search-based, LLM-based), defect detection and classification (bug triaging, duplicate detection), and automatic code repair (Automated Program Repair). Process Assurance (metrics monitoring, standards compliance) and Planning (resource allocation, risk assessment) activities together cover less than 25% of the literature.

Architecturally, multi-agent systems (MAS) prevail, where specialized agents (e.g., Generator, Validator, Reporter) collaborate via negotiation protocols or blackboard patterns. Autonomy is predominantly "moderate": the agent proposes, the human decides (human-in-the-loop), with rare cases of full autonomy in sandbox environments.

Quantified benefits include 30-70% manual effort reduction, 15-40% test coverage increase, improved defect prediction precision (F1-score >0.85). Open challenges: cross-domain generalization (overfitting on specific codebases), explainability (lack of feature attribution for LLM-based decisions), rigorous validation (absence of standardized benchmarks), and governance (accountability, bias, security). The managerial role gap reflects the absence of agents capable of strategic reasoning and organizational value alignment.