Looking At The Crystal Ball: 2025 Predictions For Agentic AI

Shailesh Manjrekar is the Chief Marketing Officer at CloudFabrix, the inventor of Robotic Data Automation Fabric and an AIOps Leader.

The crystal ball has certainly become hazy with the changes in world order, several wars brewing and the gigantic technological transformations. However, as we've done in years past, this is our attempt to bring some order to what we see happening in 2025.

2025 Will Be The Year Of Multi-Agents

Simply put, agents have the ability to ReAct (reason and act), break down tasks and execute them autonomously. Agentic AI will have levels of automation, measured as Level 1 to Level 5, depending on the level of human involvement (human-in-the-loop). Agents can sense, reason, plan, act and put LLMs (large language models) or SLMs (small language models) at the center of decision-making, which is what makes automation contextual. As I explained in a previous article on agentic AI, multi-agents are becoming a necessity to provide service as a software. Multiple agents working in tandem at the behest of an orchestrator will soon become the norm. Think of several digital workers communicating and coordinating to deliver a service.

Agentic AI Will Innovate Vertical Business Models

Experts say it is a $4.6 trillion opportunity where AI is not just eating software but salaries and services. Agentic AI is expected to innovate around business models, where licensing will be increasingly based on outcomes rather than seats to a platform or a product.

Agentic AI is set to disrupt vertical industries; however, fintech and telecommunications seem to be early adopters of agentic AI. According to a recent IDC Futurescape webinar, for every $1 invested, there is 3.7x ROI.

GenAI To Agentic AI; RAG To Agentic RAG

Will AI co-pilots and assistants be dead? Not exactly—AI co-pilots and assistants are primarily based on conversational queries and responses, with some leveraging RAG (retrieval augmented generation). They primarily provide recommendations with a human-in-the-loop.

Agents perform an action, where before responding to a conversational query, they then deliberate, contemplate, and reason before they break down the query into COTS (chain of thoughts) or GOTS (graph of thoughts) and then autonomously take action with a human-in-the-loop approach. Typically, an LLM is used to reason while an SLM (small language model) or LAM (large action model) is used for task execution. They can still intersect with each other, where co-pilots or assistants are used as UX for agents.

The Agentic AI Stack Will Continue To Emerge

The agentic AI stack can be broadly classified into the following sub-stacks: Data fabric for AI, AI agents and automation.

• Data fabric for AI: This is still going to be the heart and soul of agentic AI systems. Data source integration, data enrichment, data curation and data management will take center stage when it comes to contextualizing data for agents. Data fabric architectures with composable pipelines will become the supply chain for agents. VectorDBs have become mandatory for RAG embeddings, while GraphRAG and knowledge graphs are increasingly being used to avoid hallucinations and improve accuracy. Data fabric composable pipelines enable vectorization as well as build entity relationships with knowledge graphs.

• Agentic AI workflows, frameworks and models: More agentic frameworks will evolve. Several AI vendors (Amazon Bedrock Agents, MSFT MagneticOne, Langchain LangGraph, Salesforce AgentForce, OpenAI Swarm, Crew.ai, IBM Bee Stack) are emerging; however, each of them is putting LLMs at the center of decision-making. The orchestrator is going to be the heart and soul of multi-agent deployments—how it generates task graphs and orchestrates other tasks is going to be key to agentic workflows. Models like OpenAI O1 will be increasingly used with orchestrators and will carry the promise of AGI (artificial general intelligence) or System2 thinking. Evaluating agents is going to be a crucial building block through the ADLC (agent development lifecycle). This is because LLMs are stochastic (random) in nature unlike a deterministic piece of software.

• Agentic applications and workflows: AI-infused applications will become common. One of the promises of AI agents is to enable existing business and operational and analytical applications to leverage them. Vendors with access to domain-specific private data will become the kingmakers for developing agentic workflows and applications.

• Observability and AIOps 3.0: With the promise of autonomy comes the challenge of applying guardrails. LangChain's State of Agents survey stresses the importance of observability and AgentOps for agentic AI workflows and applications. Observability and AIOps will become extremely important in this year of agents for visibility, guardrails and operationalizing agents at scale. Privacy and security will become key inhibitors, and there will be more vendors providing audit scores for LLMs and workflows.

Movers And Shakers

Agentic AI is here and is going to be a paradigm change. Major application vendors (Microsoft, Perplexity, Salesforce, ServiceNow, etc.) are already implementing agents. Enterprise applications and intra-marketplace agents will emerge with agentic AI. Soon we will need to start discovering agents, observe telemetry from agents, provide resilience for agents and so on.

Very few operational platform vendors have adopted agentic workflows due to the challenges of dealing with massive amounts of data, data veracity and real-time telemetry data. ServiceNow and our company are a couple of the vendors dealing with operational data and already have agents to audit and approve ACL changes, open tickets when critical incidents are logged and more.

How To Prepare For Agentic AI And Its Long-Term Impact

Leaders should clearly define business objectives, use cases and the measurement KPIs for agentic workflows. They should also establish a clear operating model for these workflows with ROI and evaluate if their data applications and data platforms (business and operational) are AI-ready. Finally, they should develop responsible AI, data privacy, and security processes and engage in change management to upskill their workforces.

Forbes Technology Council is an invitation-only community for world-class CIOs, CTOs and technology executives. Do I qualify?