Beyond Chatbots: Understanding Agentic AI

Most people's mental image of AI is a chat window. You type something, you get a response. That's a useful starting point, but it's an incomplete picture of where the technology is today.

Generative AI can create things: text, images, code. Chatbots can respond to questions. Agentic AI goes a step further: It can act on your behalf. Rather than simply returning an answer, an agentic AI can execute tasks, make decisions, use external tools, and course-correct when something goes wrong, all with minimal human intervention.

This distinction is more than semantic. It represents a fundamental shift in how businesses can automate complex, multistep workflows.

At COMMON POWERUp 2026, Jesse Gorzinski, Business Architect of AI Solutions for IBM i at IBM, gave a keynote presentation about how IBM i developers and admins can use AI agents to solve problems, turbocharge productivity, and reduce overhead.

These are our top takeaways.

The Agentic AI Loop

At the core of agentic AI is a concept called the agentic loop. When given a goal, an agent doesn't just respond. It plans what it's going to do, executes that plan, then reflects on whether it succeeded. If the outcome doesn't meet the requirements, it loops back, adjusts, and tries again.

This is directly relevant to IBM i development. Imagine an agent that writes or modifies code and then attempts to compile it. If the compile fails, the agent reads the compiler output, identifies the problem, adjusts the code, and tries again, repeating until the build succeeds.

From there, it can run test cases, handle failures by revisiting the code, and ultimately deploy to a test environment. That entire software development loop — write, compile, test, deploy — can be driven autonomously by the agent.

This is one of the core reasons IBM Bob, the company's new AI assistant that's purpose-built for enterprise IBM i users, is as capable as it is. The agentic loop is built into how Bob approaches development tasks.

"IBM i is the world's first agent-native operating system," explained Gorzinski. This applies to "everything from day-to-day system activities, to work management, to security and so on."

Yesterday's Front Ends Are Today's Back Ends

One of the most useful ways to think about agentic AI is through the lens of what it replaces in a workflow. Consider a manager approving employee time-off requests. Previously, that meant logging into an HR system, navigating several screens, approving requests, then switching to email to notify employees.

Multiple front-end systems each required their own separate access.

With an agentic interface, the manager simply states the goal in natural language. The agent handles the HR system and the email, both of which become back-end systems to the agent.

What were once front ends for humans become back ends for AI.

This pattern repeats across industries. The trucking company example covered later in this post illustrates it at scale.

Guardrails and Human-in-the-Loop for Agentic AI on IBM i

Agentic AI's power comes with real risk if deployed carelessly. Several documented cases illustrate what happens when guardrails are absent or insufficient:

  • A car dealership deployed a chatbot without adequate constraints. A user negotiated the purchase of an SUV for $1, and the interaction became a legal matter.
  • Retailers deployed chatbots powered by general-purpose AI without restricting scope. Users quickly discovered they could ask the chatbots to write code or perform tasks entirely unrelated to the retailer's products.
  • A small rental company suffered the most severe outcome: An AI agent, encountering an unexpected situation, deleted the company's production database and then deleted the backup as well, all within seconds. Three months of reservation data was lost during this "vibe deletion."

These aren't arguments against agentic AI. They're arguments for building it carefully. Guardrails — rules and constraints that define the boundaries of what an agent can and cannot do — need to be part of the design from the start, not added after a problem occurs.

Equally important is the concept of human-in-the-loop: requiring human approval before the agent takes consequential actions. IBM Bob implements this by default, prompting users to approve file reads, file changes, and other operations before executing them.

Users can configure the level of oversight based on how much they trust the agent for a given workflow. The guiding principle for any agentic deployment should be "trust but verify."

How AI Agents Access Knowledge and Data: RAG

Large language models are trained on broad data, but they don't inherently know about your business, your systems, or your industry-specific documentation. Retrieval augmented generation (RAG) is the standard mechanism for bridging that gap.

With RAG, documents such as manuals, internal policies, technical references, and database schemas are processed and stored in a way that lets the agent retrieve relevant content on demand. When a user asks a question, the agent can draw on that curated knowledge base in addition to its base model training. The result is responses grounded in your actual data rather than generalizations.

For IBM i, this has been applied directly in the Bob premium package, which includes a curated library of IBM i–specific documentation, from System/36-era references through current releases. This gives the model deep platform context it wouldn't have from general training data alone.

Tools and MCP: Connecting AI to Real IBM i Systems

One of the most important concepts in agentic AI is the tool: a discrete, callable function that maps natural language intent to actual code execution. A tool has a name, a natural language description that helps the AI understand when to use it, and underlying code that does a specific, deterministic thing.

This is the key to reliability. AI models are probabilistic, meaning they're very good at approximations and generalizations. But you wouldn't trust a model to tell you your exact monthly sales figures for a CEO report based on inference alone.

A tool that queries your DB2 database directly gives you exact data every time. Agentic AI is, at its core, the merging of deterministic computing (tools, actual code, real data) with nondeterministic computing (LLM reasoning, language understanding, pattern recognition).

The protocol that makes tools interoperable across different AI systems is called Model Context Protocol (MCP). Think of it as the USB standard for AI: If you build a tool that complies with MCP, it can be plugged into any compatible AI system. You don't have to rebuild integrations for every AI platform you want to use.

IBM has released an open-source Universal MCP Server for IBM i, currently in technology preview and free to evaluate. It ships with prebuilt tools for DB2 for i, system security, work management, and other IBM i–specific functions. If you have data accessible via SQL or CL commands, you can build tools around it and expose them to any MCP-compatible AI agent.

Practical Examples of AI Agents in Action

This is how real enterprise users are getting value from agentic AI on their IBM i platforms.

Security Auditing

An agent prompted to analyze system security can query IBM i security settings through MCP tools, identify critical issues (such as public authority granted to sensitive libraries), categorize findings by severity, and generate a formatted report with remediation steps. All it takes is a single prompt.

A task that previously required attending a session, taking notes, and manually applying the findings can now be completed in minutes, with human review of the output before any action is taken.

Business Reporting

An agent connected to sales data in DB2 can retrieve, analyze, and summarize that data, identify trends and anomalies, and produce a formatted PowerPoint presentation with charts, regional breakdowns, and strategic recommendations. Again, all it takes is a single prompt.

The agent uses reasoning to determine what's interesting in the data and surfaces it without being told exactly what to look for.

Cross-System Coordination

A contractor's database of lumber inventory, supplier information, and realtime data on lead-times (retrieved via API calls embedded in DB2 queries) can be combined by an agent to produce a project plan with multiple procurement options — such as fastest delivery, fewest suppliers, lowest cost — and then present it in a formatted deck. All of this is driven by a single natural language prompt.

These examples share a common thread: The agent is doing the coordination work that previously required a human to open multiple systems, gather data manually, and synthesize it into a deliverable.

Agentic AI Skills: Encoding Complex Workflows

Beyond individual tools, agents can be given skills, or structured descriptions of how to perform multistep tasks. A skill is typically a markdown document that describes the steps, decision points, and considerations involved in a specific workflow.

Skills give agents the context needed to handle nuanced tasks correctly: what to check, what order to do things in, and what edge cases to watch for.

IBM's premium package for IBM i includes approximately 40 prebuilt skills covering common development tasks, like converting record-level access to SQL, writing RPG unit tests, compiling code, and more. Skills and tools work together: The skill describes the process; the tools execute the discrete steps.

Agentic AI Is Not Just a Chat Interface

It's worth addressing a common misconception directly: Agentic AI is not synonymous with a chat window. Chat interfaces are a convenient way to demonstrate agentic capabilities, but in production, most agentic workflows are headless, meaning they are triggered by application code, system events, scheduled processes, or user actions that don't involve a visible chat prompt at all.

On IBM i, this means an RPG program can invoke an agentic workflow. A system event in the history log can trigger an automated response. A nightly batch process can kick off an agent that analyzes data, generates a report, and emails it, all without any human in the loop at the prompt stage (though human review of output remains important).

Building agentic AI into real applications means thinking beyond the demo chat window and into how agents integrate with existing systems, processes, and automation strategies.

Real-World Example: Using Agentic AI on IBM i for Trucking Dispatch

An IBM i user in the trucking industry built a problem resolution system for drivers on the road. Previously, a driver with a breakdown would call the front office, and someone in that office would manually coordinate with repair facilities, parts suppliers, and tow companies.

Each interaction required access to a separate system. It was a long, cumbersome process.

The company replaced that workflow with an agentic application. A driver reports the issue through an app. The agent, connected to repair scheduling systems, parts supplier APIs, and tow dispatch systems as back-end services, then coordinates the full resolution automatically. What previously took hours of hold time and manual coordination now resolves in minutes.

The business impact extends beyond driver satisfaction. Faster resolution means less idle time, which directly affects driver pay. Predictable resolution also reduces the risk premium that trucking companies build into bids, potentially improving competitiveness and revenue.

All of this company's data, including their transportation management system and operational data, runs on IBM i.

The throughline across all of this is the same: Agentic AI works by connecting powerful reasoning capabilities to real data and real systems through well-defined tools, governed by appropriate guardrails, with humans retaining oversight of the outputs that matter.

For IBM i shops, the infrastructure to participate in this shift is available today with CloudFirst. Get in touch today.