AIM, in partnership with NVIDIA, not too long ago concluded an Agentic AI workshop. The session, titled ‘From Scratch to Answer: Multi-Agent AI for Advanced Duties’, was led by Shreyans Dhankhar, senior answer architect at NVIDIA. It provided contributors a sensible understanding of constructing subtle AI programs able to autonomous activity execution.
Dhankhar stated that brokers are essentially “software program applications that may autonomously execute duties for customers or different functions”. The core element enabling this autonomy is the massive language mannequin (LLM), possessing capabilities for pondering, reflection, and power utilisation.
He highlighted brokers’ benefit in balancing latency and price for enhanced activity effectiveness, notably in eventualities demanding complexity, reliability, and consistency.
Sharing an instance, he stated that brokers can work together with knowledge in pure language and may generate SQL queries, showcasing they empower non-experts to work together with knowledge.
Constructing Blocks of Agentic Methods
The workshop progressed from foundational ideas to sensible implementation. Dhankhar outlined totally different agentic patterns—beginning with single-hop brokers for easy duties to multi-actor chain brokers, the place the output of 1 agent sequentially feeds into one other, and hierarchical agent frameworks, that includes a supervisor agent delegating duties to sub-agents. Furthermore, the mixing of human-in-the-loop capabilities was mentioned, notably in refining artificial knowledge technology.
A good portion of the session was devoted to a hands-on demonstration of constructing a easy ReAct (Motive, Act, Observe, Suppose) agent from scratch utilizing fundamental Python. This train illustrated the core ideas of agent operation: fascinated by an issue, taking an motion (utilizing an outlined software or operate), observing the outcome, after which iterating. The instance used a reminiscence mannequin calculator to find out the video random entry reminiscence (VRAM) requirement for an LLM primarily based on its measurement and precision.
Dhankhar underscored the significance of a well-defined immediate, performing because the “system immediate or system message” to information the LLM’s behaviour.
The workshop showcased a multi-agent system instance, simulating an e-commerce workflow. This concerned a number of brokers with particular functionalities—liable for duties like climate verify, product search, and order processing—speaking sequentially to fulfil a person request. This demonstration highlighted message-based communication and stateful dialog administration inside a multi-agent structure.
Superior Functions: Artificial Knowledge Era
The ultimate hands-on train centered on constructing a system for supervised fine-tuning knowledge set technology. Dhankhar introduced a workflow the place totally different brokers collaborate to course of content material from an internet site, extract key ideas and relationships, generate questions, and design a schema for the info. An orchestrator agent managed the interplay between these specialised brokers.
“That is some pipeline that may make it easier to out in that,” Dhankhar defined, emphasising the relevance of artificial knowledge technology in eventualities the place adequate real-world knowledge for fine-tuning is missing. The demonstration concerned processing a dummy web site and producing question-answer pairs in a JSONL format, a standard customary for coaching knowledge. He additionally confirmed how the system might adapt to totally different enter varieties, similar to plain textual content recordsdata.
Important Concerns and Sources
Dhankhar concluded by outlining essential necessities for sturdy agentic programs, together with parallel thread execution, scalability, load balancing, error dealing with, monitoring, model management, and safety. He additionally supplied an outline of the agentic AI ecosystem, mentioning orchestration frameworks like CrewAI, LlamaIndex, Autogen, and LangChain, in addition to basis fashions and vector databases.
Attendees had been launched to NVIDIA AI-Q, a toolkit for constructing and orchestrating brokers throughout a number of frameworks. Dhankhar additionally highlighted free self-paced NVIDIA Deep Studying Institute (DLI) programs, /providing hands-on expertise in numerous AI domains.
Click on right here to view and register for the programs.
He additional pointed to NVIDIA developer instruments and sources, together with NVIDIA Inference Microservice (NIM) and NeMo, and inspired contributors to discover the on-demand classes from GTC 2025 for deeper insights into AI.
Within the Q&A session, Dhankhar addressed questions starting from GPU necessities for giant fashions to dealing with conflicting agent targets and mitigating hallucinations. He careworn the context-dependent nature of many options and inspired experimentation with the supplied code and sources. The workshop served as a helpful introduction to the sensible facets of constructing multi-agent AI programs, empowering attendees to discover this transformative know-how additional.
The put up NVIDIA Demonstrates Actual-World Functions of Agentic AI in Joint Workshop with AIM appeared first on Analytics India Journal.