IBM and Agentic AI: Transforming Businesses

The Importance of AI Agents in the Evolution of Businesses

Guides have assisted humanity throughout its history. Prehistoric civilizations understood that the sun and the moon could be used to navigate vast distances, whether on land or at sea. Over time, various voyages facilitated the production of maps for better planning and faster travel times to repeated destinations. Centuries later, the introduction of the compass allowed sailors to achieve greater accuracy in seeking unexplored destinations. Today, GPS navigation applications guide each of our movements. In the current world of agentic AI, AI agents undoubtedly have the potential to enable an evolutionary adoption of AI, transforming industries as we know them. However, an intelligent guide, agentic logic, is necessary to realize this potential by fostering high agent quality, cost-effectiveness, and, consequently, end-user trust.

Challenges and Solutions in Business Workflows

Numerous studies have highlighted the overwhelming failure of AI pilots, while others have also emphasized the need for AI to operate at the heart of business workflows to enable evolutionary adoption. To better understand this phenomenon and the associated claim, an analysis of business workflows is necessary. These workflows are:

• Dynamic and long-lasting
• Possess a plethora of APIs, databases, and services
• Often constrained by business policies and/or regulations

For an agent to function effectively, given these characteristics, it is naturally necessary to expand the context of the model, which leading LLMs certainly possess, but at what cost? Increased hallucinations, token consumption? Moreover, can LLMs be equipped with an intelligent guide, a GPS, to enable the execution of agentic AI at the heart of the workflow, leading to more desirable outcomes? We tested these hypotheses by designing and building agents, equipped with relevant agent logic, for IBM's offerings while fully considering the above characteristics. These offerings pertain to some of the most challenging tasks faced by subject matter experts managing various stages of the enterprise software delivery cycle for critical workloads, including:

• Understanding applications written in legacy code (Cobol / PL/1)
• Accelerating test generation for developers
• Proactively responding to incidents and improving application resilience
• Automating compliance modernization for critical environments

Before examining each of these areas in detail, let’s define what characterizes agent logic. Agent logic consists of software primitives, such as knowledge graphs, algorithms, and program analysis libraries, that operate at the agentic level (within an agent framework) and can intentionally guide the LLM in the direction of the business workflow, thereby reducing the context space. In doing so, it tends to generate higher-performing results more cost-effectively. Let’s now explore how agent logic can achieve such results in each of the four areas mentioned above.

Understanding Applications Written in Legacy Code (Cobol / PL/1) - Program Analysis

IBM watsonx Code assistant for Z (WCA4Z), used to accelerate the development and modernization of mainframe applications with AI and automation, is equipped with an App Insights agent for application understanding—one of the primary areas of interest for enterprise clients managing critical workloads on IBM mainframes. This agent leverages deep static analysis of the application and stores a pre-indexed representation in a database schema that spans hundreds of interconnected tables with complex semantics, allowing the agent to retrieve accurate and structured information already available; thus improving the accuracy of responses, reducing token usage, and minimizing repeated interactions with the language model (Mistral Medium 250B in this case). This approach, when applied to multiple critical legacy systems (up to 1 million lines of code and 1,000 programs), maintains marginally superior application understanding performance with token consumption approximately 30 times lower than that of a leading LLM-only approach.

Accelerating Test Generation for Developers with Aster - Program Analysis

Aster is a proprietary program analysis and data pre- and post-processing library from IBM used for agent-based generation of unit, integration, API, and change tests; which, based on the analysis of multiple developer communities, achieves superior ratings compared to various open-source tools or tests written by developers. Based on these results and superior benchmarks for line, branch, and method coverage compared to similar open-source tools (integration tests) and zero-shot LLM and coding agents (unit tests), all tested on open-source applications, we ran Aster in pre-production mode on over 75 Java applications from IBM CIO (up to 560 classes and 67,000 lines of code) with the Devstral 24B model. The stable results to date show a 20% to 45% improvement in line, branch, and method coverage, coupled with superior performance on a subset of these applications compared to a leading coding agent with several orders of magnitude lower token consumption (up to 15 times). The reason for these results is that the output of the program analysis (used to prompt and "focus" the LLM) combined with sub-agents to increase coverage and remedy execution and compilation errors yields a higher-performing result with a significant reduction in costs.

Proactive Incident Response and Application Resilience Improvement - Knowledge Graphs, Program Analysis Libraries, and Observability-Driven Orchestration

While the context of LLMs for application-related use cases, as described in points 1 and 2, is "restricted" to the application's source code, for real-time management of applications on the deployed infrastructure, the entire computing stack comes into play. Here, we define a knowledge graph (KG) encompassing entities (microservices, database/middleware services, MELT, etc.) coupled with embedded ("tribal") knowledge from domain experts. With such a graph and limiting the LLM to local reasoning for non-deterministic results, an observability-driven approach is used to reduce the context space encompassing the computing stack and the underlying application source code (if relevant) for root cause analysis of incidents (and other use cases). With this approach, leveraging Instana's equivalent data model, we found that the proprietary Instana agent "I3" (Intelligent Incident Investigation) achieved up to 4.0 times improvement over the ReAct agent with GPT-5.1, measured using ITBench. With Gemini 3 Flash, the performance of the ReAct agent improves to be 17% below that of the I3 agent while consuming 1.6 times more tokens. We extended this approach to the source code with agents for code analysis (leveraging program dependency graphs) and bug remediation (leveraging inference scale), also tested on ITBench, illustrating superior performance for code analysis and bug remediation agents (Gemini 2.5 Flash) compared to a leading coding agent, both for identifying the responsible microservice (3.0 times) and for bug fixing (1.6 times), while consuming 3.7 times and 5.9 times fewer tokens, respectively. This multi-agent system was announced at IBM Think as part of the new IBM Concert platform for shift-left IT operations and is also being tested internally with IBM CIO.

Automating IT Compliance Modernization for Critical Environments - Algorithms and Adaptive Planning and Orchestration

Businesses face increasingly complex and fragmented compliance requirements, forcing teams to spend considerable time manually creating controls, assessments, and remediation plans. No centralized knowledge exists, and fixes are written manually, introducing a risk of errors and security vulnerabilities. Given that compliance work is complex and involves multiple steps, it requires coordinated automation driven by policies across specialized agents rather than manual effort or simple AI prompts. Our multi-agent system automates compliance by algorithmically breaking down complex tasks into coordinated steps, using adaptive planning, dynamic decomposition, and workflow sequencing with continuous feedback to iteratively identify fixes and expand assessments. It is 1.3 to 2.0 times more effective than previous agents (Claude 4 Sonnet) using fixed planning strategies, as also measured using ITBench. This approach transforms compliance into a continuously guided self-correcting process and significantly improves outcomes, especially in complex scenarios, increasing success rates from a few percentage points to over 80% (Claude 4 Sonnet). This multi-agent system and over 16,000 digitized control mappings were unveiled as part of IBM Sovereign Core at IBM Think, integrated with monitoring, drift detection, providing automatic evidence generation, ensuring that audit evidence remains under the client's secure control.

The examples above illustrate the impact of agent logic in reducing the context of LLMs and in guiding the LLM to traverse the heart of the workflow in a highly performant and cost-effective manner. Additionally, we have employed similar approaches in two case studies, one with a configurable generalist agent (CUGA) in the healthcare domain and the other for condition-based maintenance of physical assets with IBM Global Real Estate.

Case Studies

Case Study 1: Configurable Generalist Agent (CUGA) - Health Benchmark - Application of Algorithmic Policies

The following example from the customer service of a health insurance company succinctly illustrates why agentic systems outperform conversational models solely based on LLMs in regulated environments. The CUGA policy system (configurable generalist agent) implements policy as code for agent governance, which is applied in real-time regardless of model prompts and without fine-tuning. Our experiments show that the agent's policy system fills significant gaps in task correction by applying structured workflows and safe processing of intents.