By: Sunil Agrawal, Associate Vice President, LTIMindtree

Back in the day, artificial intelligence (AI) was deployed over a single centralized system that did everything from processing data to executing commands This approach was feasible then, as AI’s capabilities were limited to basic natural language processing and reasoning. However, with advancements in generative reinforcement learning and other algorithms, we are witnessing the adoption of AI models that function as independent agents with specific roles and responsibilities, capable of dynamic adaptation to user inputs and environments, better known as Agentic AI. It is being scaled and deployed into different application areas in a multi-agent manner. Large language models (LLMs) and other generative tools have become powerful enough to execute multiple tasks simultaneously. However, their contextual capabilities are inversely proportional to the width of the application domain it must operate on.

To tackle this issue, organizations are looking at reducing the domain of application for these powerful AI models, clearly defining their set of responsibilities, and establishing requirements in a specific context. These compact and contextual models are termed AI agents or agentic AI.

The release of AutoGPT and BabyAGI in 2023 marked the first practical instance of an AI agent. BabyAGI works on the premise of using generative AI to automate as many simple tasks as possible. The key feature of such AI agents is their ability to dynamically adapt to user inputs and their working environments. These agents are like assistants, capable of generating content independently – but still require human guidance for execution and result validation in complex application areas. They can enhance user capabilities to get through repetitive tasks quickly.

Here are some important questions: how do these AI agents operate? Can we use them in a use case, and can we trust an agent enough to autonomously take a task, execute it, and give us results? In this blog, we attempt to answer them.

The Anatomy of AI Agents: Enhancing Planning and Reasoning with Generative Models and CoT Techniques

An AI agent comprises a generative model for interpreting inputs, a planning and reasoning aspect that imparts dynamicity to the agent, appropriate tool calling, integrations, a reflection loop, and a policy and governance layer. Users must plan tasks and establish a set of responsibilities and sequence of execution. Once an agent understands the intent and the plan, it collects information from the database, selects the appropriate tool or API integration, executes the code sequence as per the assigned responsibilities, and generates the output, which is then sent through a reflection loop and tested for accuracy. The generated data is compared against the ideal output generated by the discriminator model and, if required, sent back to the agent for refinement. The final output is then sent through the governance function.

Though the fundamentals of agentic workflow exist, current generative models don’t offer the level of planning and reasoning capabilities required to make AI agents work. Some models have incorporated chain-of-thoughts (CoT) reasoning and reinforcement learning, where the user guides the model through every step of reasoning by asking questions and giving prompts to reflect on the output. This ensures that the model is self-correcting and learning throughout the process. The latest GPT model has integrated CoT, making it think at each step of prompting. While the end goal hasn’t been attained yet, this progress is noteworthy.

Future-proofing Software Development: The Power of AI Agents in Managing Roles and Responsibilities

Software engineering is modular with object-oriented approaches – with programmers, front-end and back-end coders, unit testers, project managers, and respective toolsets for development – with different sets of responsibilities and executing different tasks. While leveraging a single powerful large model to automate the entire cycle is possible, setting it up is a huge challenge.

The adoption of agentic AI allows us to represent these responsibilities as agent nodes, with each role becoming an independent agent. We must program the agent according to its job description, its intended roles and responsibilities, and intended and expected outputs. Once done, we can deploy and integrate them using specific APIs for each agent. The agents will establish a connection between the nodes over multiple iterations.

Once these connections are established, we can replace and upgrade the nodes without integration. This is because the generative model handles the connection between the intended function and a specific API call. We can specifically program agents to check the below-expectation outputs and trigger a command to refine them.  This makes the pipeline robust and future-proof, and reduces hallucinations.

The Trust Dilemma: Why Current AI Agents Can’t Operate Autonomously Yet

So far, we have discussed how AI agents operate and their use cases. Now, the only question remains: can we trust an agent enough to autonomously take up a task, execute it, and give us results?

At present, we cannot trust current AI agents with this level of autonomy. We have reached this conclusion since existing generative models have limited contextual understanding, planning, and reasoning capabilities. The level of complexity required to connect and synchronize with every individual AI agent is very high. Mapping structured thinking on an AI agent will produce larger agent packages, which will incur a heavy load on infrastructure, eventually increasing OPEX.

While there is no guarantee that these AI agents will not hallucinate, there may be some instances that can be gauged, measured, and reduced,  These agents are like black boxes – We have the tools and capabilities to tinker with their logic and reasoning, but it is extremely difficult to know precisely how these models will think.

The Future of AI: What You Need to Know?

It is safe to say that AI agents will run on generative models and make their own decisions, which is better than off-the-shelf model-based agents that guess the output based on reasoning and training on non-contextual data. Thus, the next logical step to get started is to create a domain-specific agent for simple and repetitive tasks – Track the interaction between the agent and its environment and based on the feedback, upgrade or train it. It is cheaper and easier to rectify and retrain straightforward agents rather than complex ones that are deployed at critical applications. With iterations, human-centricity, AI ethics, and governance considerations can be built in.

While we have tried to answer some important questions, there are many more that need to be answered including the global implications of such systems, how they can be made human-centric, will they work in practical, real-world situations, and what will we gain out of them. We will try and answer these in the next blog.