Not all AI agents work the same way. Some just react to what's in front of them. Others plan ahead, weigh tradeoffs, or improve from experience. Understanding the four core types, reactive, goal-based, utility-based, and learning agents, tells you what an agent can actually handle before you deploy one.
This isn't about Zamp HR or payroll software, and it isn't about the zamp.com sales-tax platform. This is about AI agents, the systems that perceive, decide, and act, and the different architectures behind them.
A reactive agent responds directly to its current input with no memory of the past and no model of what might happen next. It follows a fixed set of condition-action rules: if X happens, do Y.
Example: A fraud-detection agent that flags any transaction over $10,000 from a new vendor for review. It doesn't remember prior transactions or model the vendor's future behavior. It just checks the current transaction against a rule and acts.
Reactive agents are fast and predictable, which makes them useful for narrow, well-defined tasks. They break down fast outside that narrow scope, since they have no way to reason about context they haven't been explicitly told to check.
A goal-based agent (also called a model-based agent) keeps an internal model of the world and picks actions based on whether they move it closer to a defined goal. It can consider several possible action sequences and choose the one that gets it there.
Example: An accounts payable agent tasked with closing the books by month-end. It doesn't just process invoices one at a time; it tracks which vendors are outstanding, which approvals are pending, and sequences its own work (chase approvals first, batch payments second) to hit the deadline.
This is the shift from "respond to what's in front of me" to "work toward an outcome." Most of what enterprises call AI agents today, in accounts payable, procurement, or customer support, are goal-based, because the work involves multi-step sequencing toward a defined end state.
A utility-based agent goes further than "reach the goal." It ranks multiple ways of reaching the goal by how good each one is, using a utility function, and picks the best one. This matters when there are tradeoffs: speed versus cost, accuracy versus coverage.
Example: A vendor-onboarding agent that could route a new supplier through express verification (fast, slightly higher risk) or full compliance review (slower, lower risk). A utility-based agent weighs the vendor's transaction size and industry risk profile and picks the path with the best expected outcome, not just the fastest one.
Utility-based agents are where "AI agent" starts to mean something closer to a judgment call rather than a fixed procedure.
A learning agent improves its own performance over time based on feedback from its environment. It doesn't just execute a fixed model; it updates that model as it collects more data.
Example: A customer-support agent that starts with a baseline set of resolution paths and, over thousands of tickets, learns which responses actually resolve issues on the first reply versus which ones generate a follow-up. Its routing improves without a human rewriting its rules.
Learning agents carry the most long-term value and the most operational risk, since their behavior shifts over time and needs monitoring, not just initial testing.
Most production AI agents aren't purely one type. A single AI agent deployed for accounts payable is typically goal-based for sequencing work and utility-based for deciding which invoices need human review versus straight-through processing. See real enterprise use cases for what this looks like across functions, or start with what an AI agent actually is if you're still on the definitional question.
The type isn't a marketing label, it's a description of how the agent decides what to do next. That distinction matters when you're evaluating AI agent platforms: a vendor claiming "AI agents" that only does reactive rule-matching is a very different product from one running utility-based or learning agents against real business outcomes.
What are the 4 types of AI agents? Reactive agents (respond to current input only), goal-based agents (plan toward a defined outcome), utility-based agents (rank options by expected value), and learning agents (improve from feedback over time).
Is a chatbot a type of AI agent? A basic scripted chatbot is closer to a reactive agent. A chatbot that plans multi-turn conversations toward resolving an issue and adapts based on outcomes crosses into goal-based or learning-agent territory.
What type of agent is an AI employee? Most AI employees combine goal-based sequencing (getting a full workflow done) with utility-based judgment (deciding which cases need a human) and often learning components that improve over time.
