The agentic future of business: from voice assistants to enterprise intelligence

The Agentic Web: Building Reliable, Scalable AI for the Future of Business

Babak Hodjat of Cognizant opened by tracing AI’s cyclical journey. Early attempts in the 1970s and 80s to build general intelligence failed due to limited compute and algorithms, leading to specialized “agents” for contained environments. The 1990s saw multi-agent systems emerge, evolving into distributed AI where emergent behavior from simpler agents solved complex problems. Deep learning, with its basic activation functions, represents an extreme form, with system knowledge embedded in network weights.

Today’s AI is dominated by transformer-based large language models (LLMs), excelling in language understanding, translation, reasoning, and basic math. While initially seen as universal, LLMs are most effective when “contained” within an agent, given specific responsibilities, tools, and connections to other agents, marking a return to multi-agency. Mr. Hodjat clarified that an LLM generates code, but an agent can *run* and iterate on it within a container, leading to superior performance. This ability to “do stuff” is the key distinction, forming multi-agent systems that break down organizational and software silos.

Cognizant’s open-source Neurosun accelerator demonstrates practical multi-agent applications. Examples include a telco system where agents on network nodes autonomously manage load balancing and report up a hierarchy for consolidated insights. Similar hierarchical structures are applied in farming, with local agents making daily operational decisions and reporting regionally. Multi-agent systems also streamline the software development life cycle, automating tasks like test case generation and code reviews. These systems can be interconnected for comprehensive processes.

Furthermore, agents can facilitate inter-business interactions, as shown by a conceptual system where agents representing Airbnb, Expedia, and Booking.com negotiate to find optimal travel options for a consumer. To ensure reliability, agentic systems can monitor other agents, scrutinizing logs, decisions, costs, and response times, flagging issues or involving human intervention. Mr. Hodjat also presented “vibe coding,” a method to dynamically construct agent networks from natural language descriptions, automatically defining agents, their interconnections, and example queries.

Addressing concerns about long reasoning chains, Mr. Hodjat highlighted a recent Cognizant paper. It demonstrates that error correction techniques enable multi-agent systems to reliably solve complex problems, like the Tower of Hanoi, with thousands or millions of agents without error. This contrasts with previous findings of catastrophic breakdowns in single-model chain-of-thought reasoning. The paper posits a future “agentic web” where interconnected agents augment every organizational component, ensuring reliability. Mr. Hodjat concluded by inviting participation in the open-source Neurosun platform.