Indian IT Builds the Missing Layers in Open-Source Agentic AI Frameworks

A recent academic review warning that most agentic AI frameworks remain immature has drawn detailed responses from Indian IT services firms building such systems for enterprise deployment.

The research paper published in IEEE Software, titled ‘Agentic AI Frameworks Under the Microscope: What Works, What Doesn’t’, evaluated widely used open-source platforms designed to build AI agents capable of perceiving context, reasoning and acting autonomously.

Authored by Karthik Vaidhyanathan, an Assistant Professor at the Software Engineering Research Center, IIIT Hyderabad, and Davide Taibi of the University of Southern Denmark, the study examined 243 GitHub repositories with more than 5,000 stars before narrowing the field to 10 general-purpose frameworks. These included AutoGPT, Microsoft Semantic Kernel, LangGraph, CrewAI, AutoGen and Google’s Agent Development Kit.

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

The key shortcomings identified included constrained memory management—often limited to short-term or vector-store implementations—inconsistent observability and underdeveloped security controls. The paper also cautioned that popularity metrics, such as GitHub stars, are reliable measures of production readiness.

Enterprise Layering Beyond Base Frameworks

Responding to the findings, Tech Mahindra and Persistent Systems said their production platforms are built on open frameworks but significantly extended for enterprise use.

“Our agentic platform, Orion, is built on widely adopted open-source frameworks such as CrewAI and LangGraph, with additional frameworks already embedded and others on our roadmap,” Nikhil Malhotra, Chief Innovation Officer and Global Head of AI, Tech Mahindra, told AIM.

“We have deliberately treated these frameworks as foundational building blocks rather than end solutions and have layered significant enterprise-grade capabilities on top.”

According to Malhotra, Orion incorporates a dedicated agent evaluation framework, visibility into ‘agent thought trajectories’ to explain how decision pathways, robust retrieval-augmented generation (RAG) pipelines and a supervised fine-tuning pipeline for small language models.

The platform also includes a business process integration layer that exposes its capabilities as APIs, enabling workflows to be embedded directly into enterprise systems.

Persistent Systems has taken a similar approach with its internal platform, AssIstX, combining open frameworks with proprietary enterprise components.

“AssIstX is intentionally designed as a hybrid platform. We leverage proven foundational frameworks where they add value, while building proprietary orchestration, governance and enterprise-integration layers that reflect the realities of large, complex organisations,” Debashis Singh, Chief Information Officer, Persistent Systems, told AIM. 

At its foundation, Persistent uses Microsoft Semantic Kernel to structure agent behaviour and LangGraph to manage workflow execution, memory and governance across multi-agent workflows.

“Where AssIstX differentiates itself is in the enterprise layers built on top of these foundational frameworks,” Singh adds.

Strengthening Memory and Workflow Continuity

The academic review highlighted memory limitations as a recurring weakness in current frameworks.

Tech Mahindra said Orion addresses this by separating short-term, long-term and contextual memory for each agent and workflow.

“Users are given direct control over this through simple configuration options, including the ability to enable or disable memory persistence as required,” Malhotra explains.

“When enabled, Orion securely retains relevant context from prior and ongoing interactions, as well as other useful signals, so that agents can operate with continuity across long-running or multi-session workflows.”

Persistent described AssIstX as being designed for ongoing enterprise processes rather than isolated prompt-based interactions.

“AssIstX is designed around the reality that enterprise work is continuous, not transactional,” Singh said.

“Rather than relying on isolated retrieval, AssIstX carries forward workflow state, role-based context and system-derived attributes such as location, eligibility and authorisation.”

Human Oversight and Transparent Reasoning

The study also underscored the importance of reasoning transparency and human-in-the-loop safeguards as organisations move from deterministic workflows to probabilistic, agent-driven systems.

Tech Mahindra said Orion follows a ‘Model-as-a-Service’ design principle, allowing organisations to configure any large or small language model within the platform.

“For high-risk or regulated decisions, Orion does not impose rigid workflow constraints. Instead, it gives users complete autonomy to design agentic workflows that explicitly include human-in-the-loop checkpoints wherever required,” Malhotra says.

Persistent also emphasised structured human oversight.

“Fundamentally, the system is designed to remain human-led,” Singh said.

“For regulated, high-impact and ambiguous scenarios, AssIstX either seeks confirmation, escalates to a human reviewer, or routes the request through a governed alternative workflow.”

Observability, Audit and Security

Observability gaps were another concern raised in the paper.

Tech Mahindra said Orion this through comprehensive logging, alerting and monitoring mechanisms that span the entire agent lifecycle.

A key capability, according to Malhotra, is the platform’s ‘agent trajectory’ feature. This captures and records the sequence of steps and reasoning an agent follows before reaching a decision. The records are retained to support traceability and auditing in live production environments.

Persistent Systems said governance and security are embedded into AssIstX by design.

“Security and governance are embedded into AssIstX by design,” Singh said, adding that access is governed through enterprise identity management, role-based and user-based controls, and persona-aware data entitlements.

According to him, guardrails extend to prompt-level controls, context-aware responses and continuous logging and monitoring. These controls are aligned with standards such as ISO 42001, ISO 27001 and SOC 2.

From Experimental to Operational

The academic paper described itself as a diagnostic guide rather than an endorsement of any single framework. It urged practitioners to prioritise memory persistence, reasoning transparency and operational visibility.

While the underlying open-source ecosystem may remain uneven, Indian IT firms argue that enterprise deployments depend less on the maturity of base frameworks and more on how governance, integration and oversight layers are engineered on top of them.

Tech Mahindra said all Orion modules are deployed in live production environments and architected to scale on Kubernetes. Persistent said AssIstX is fully operational across the company, supporting over 26,500 employees.

Notably, Tier 1 and other mid-cap Indian IT companies did not respond to queries by AIM.