Out of 10 Popular Agentic AI Frameworks, 0 are Foolproof, Finds IIIT Hyderabad Prof

A systematic review of agentic AI frameworks has found that despite rapid innovation and growing adoption, most platforms remain immature and are not yet ready for reliable, large-scale use.

The findings are detailed in a research paper published in IEEE Software by Karthik Vaidhyanathan, assistant professor at IIIT Hyderabad; and Davide Taibi, chair of software architecture at the University of Southern Denmark. 

The study titled ‘Agentic AI Frameworks Under the Microscope - What Works, What Doesn’t’ examined the current landscape of agentic AI frameworks and software tools used to build AI agents that can perceive context, make decisions, and act autonomously.

The researchers conducted a structured search on GitHub, filtering repositories with more than 5,000 stars. This initial screening identified 243 projects. After applying additional inclusion criteria, focusing on general-purpose frameworks and excluding narrow tools or wrappers, the list was narrowed to 10 representative agentic AI frameworks.

The final set includes AutoGPT (uses Python language), Microsoft Semantic Kernel (Python, C#, Java), LangGraph (Python, JavaScript), Agno (Python), MetaGPT (Python), CrewAI (Python), AgentGPT (JavaScript/TypeScript), SmolAgents (Python), AutoGen (Python), and Google’s Agent Development Kit (Python).

The authors emphasised that popularity alone was not sufficient. “Our review of the 10 most representative general-purpose agentic AI frameworks reveals a maturing but fragmented landscape,” they stated in the paper, adding that broad feature sets often conceal limited depth.

Each framework was evaluated across five dimensions: goal handling and interaction, planning and reasoning, architectural patterns, state and memory management, and community and ecosystem support. These dimensions were derived from existing academic literature and observable characteristics in open-source projects, which the authors describe as “core principles of agentic system design.”

The review identified several common trends. Proactive goal processing—where agents infer next steps without explicit prompts—has become widespread. Also, multipath planning is increasingly replacing linear task execution, and multi-agent collaboration is now more common than single-agent orchestration.

Weaknesses 

At the same time, the study highlighted persistent weaknesses. Memory management remains limited, with many frameworks relying primarily on short-term context or vector stores. 

“Memory management is often limited to short-term or vector-store implementations,” the authors noted, constraining an agent’s ability to maintain continuity across long or complex workflows.

Observability and security are also uneven. Only a small number of frameworks provide advanced tracing and telemetry, while others offer little more than basic task logs. Security features are described as “largely underdeveloped,” with safeguards typically limited to mechanisms such as sandboxed execution or policy hooks.

These gaps are significant because agentic AI represents “a change from fully deterministic preprogrammed workflows to more probabilistic adaptive systems that can handle uncertainty and evolving requirements,” the paper noted. Without strong observability, error handling and governance, such systems are difficult to trust in production environments.

The study also found that community adoption metrics, such as GitHub stars, do not reliably indicate production readiness. Documentation quality, contributor diversity, deployment options and language support vary widely across frameworks, reinforcing what the authors characterised as an ‘uneven ecosystem’.

Rather than recommending a single platform, the paper positioned itself as a diagnostic guide. Practitioners, the authors argued, should focus less on claims of autonomy and more on fundamentals such as memory persistence, reasoning transparency, human-in-the-loop controls and operational visibility.

The paper concluded that “most agentic AI frameworks are still immature, with limited memory, basic planning, and uneven ecosystem support,” and that “significant improvements are needed before these frameworks can enable reliable and large-scale adoption in practice.”

Speaking to AIM, Vaidhyanathan said the frameworks represent the most widely used agentic AI platforms, adding that organisations developing agentic systems are likely to rely on at least one of them due to their popularity.

“These frameworks are indicative of the most available and popular frameworks,” he said. “Among them, Google ADK and Semantic Kernel are more enterprise-ready as well.”