What is Agentic AI? The shift to autonomous systems
Agentic AI refers to autonomous software systems powered by foundational models that can independently reason, break down complex tasks, use external tools, and execute multi-step workflows to achieve open-ended goals without continuous human prompting. Unlike traditional static chatbots that rely on immediate, reactive input-output cycles, agentic systems use persistent memory, self-reflection loops, and deterministic orchestrations to manage entire processes from initiation to verification.
To make it easier to understand, here is an example and analogy of agentic AI in real-world use cases.
An analyst types one instruction: "Find out why our distribution costs rose 12% last quarter."
The system breaks that into steps, writes a database query, and runs it. The query fails – a column was renamed during a migration. It reads the error, inspects the schema, rewrites the query, and tries again. This time it works. Shipping volume is flat, so the cause is pricing. It searches the company's documents, finds the carrier contract, and locates a fuel surcharge that activated in April.
Nobody typed anything between those steps.
That loop – decide, act, check the result, adjust – is what makes a system agentic. Generative AI produces content when you prompt it, and you hold the plan. Agentic AI holds the plan itself: decomposing goals, choosing tools, and changing course when something fails.
This isn't the first attempt at automating multi-step work. Robotic process automation has done it for years by following scripts written in advance, and research on its adoption in accounting firms found the recurring problem was brittleness – scripts held until a form changed or an exception appeared that nobody had anticipated (Perdana, Lee & Chu, IJAIS, 2023). Agentic systems trade brittleness for flexibility. They can handle the exception. They can also invent a response to it that nobody sanctioned.
Generative AI vs. Agentic AI, what's the difference?
Generative AI creates content. Agentic AI executes goals. This simple distinction marks the next major shift in artificial intelligence. Moving from standard Generative AI to Agentic AI represents a massive leap in technical capability. Standard GenAI is fundamentally stateless and prompt-dependent; it requires a human to act as the primary orchestrator, reviewing outputs and providing subsequent instructions.
Agentic systems shift the human out of the immediate operational loop and into a governance role. By using advanced architectural frameworks like "reason-and-action" (ReAct) loops, these agents evaluate their own intermediate steps, recognise failures, and dynamically pivot their strategies to achieve the target objective.
| Capability dimension | Generative AI | Agentic AI |
|---|---|---|
| Operational workflow | Singular, prompt-driven execution. | Automated task decomposition and multi-step execution. |
| Tool utilisation | Restricted to native model interfaces. | Autonomous API tool-chaining, database querying, and application control. |
| Memory profiles | Short-term context windows (stateless). | Vector stores, episodic state preservation, and semantic archives. |
| Error handling | Hallucination propagation unless reprompted. | Automated self-reflection, debugging loops, and consensus validation. |
How it works
Three capabilities make the loop possible. The system plans: it decomposes a goal into ordered steps and decides what finished looks like before it starts. It uses tools to reach the world outside itself – database connections, code execution, web browsers, internal APIs – choosing when to call each one and folding the result back into its next decision. And it keeps memory: a working scratchpad for the current task, plus stored context that persists across sessions.
None of that is remarkable in isolation. What makes it powerful, and risky, is the middle one. Tools are the point at which a text generator becomes something that can change the state of your business.
Three things that break
The model isn't tracking the world the way you assume. An agent's premise is that it holds an accurate picture of things and updates it as it acts. Mechanistic research finds that language models don't build that picture up step by step as events unfold – they assemble it at the moment the question is finally asked (Tang et al., ICML 2026).
Fluent is not the same as correct. Agentic systems are overwhelmingly built and benchmarked in English, and they degrade outside it in ways accuracy scores don't catch. An agent writing to an Indonesian customer, a government office, or a senior colleague has to get register right – a requirement examined directly in Javanese (Farhansyah et al., ACL, 2025). Fluent and socially wrong is still a failure.
Nobody has settled how to grade the work. A chatbot's output gets read by a human immediately. An agent's twelve intermediate steps do not. The instinct is to have a second model mark the first one's homework, but an automatic metric only means something once it has been calibrated against what people actually judge to be good – and that calibration is itself an open research problem (Winata et al., MetaMetrics, ICLR 2025). If you can't measure it, you can't deploy it responsibly.
What this work actually requires
Not prompt writing. The hard questions are how you evaluate a process no human reads end to end, how you build and validate the data pipelines feeding it, and who approves what when it goes wrong. Those are statistics and machine learning, data engineering, and governance problems – and they don't expire when the current framework does.
They are also being worked out here in Indonesia, not imported. Research on algorithmic trust and regulation for Indonesia's central banking sector, supported in part by Bank Indonesia, has been mapping what governance of automated financial decision-making should look like in this jurisdiction (Perdana, Arifin & Quadrianto, Technology in Society, 2025). Anyone deploying an agentic system inside an Indonesian bank, hospital, or ministry will be building against specifications like these.
Take the next step in your AI & data science career
Understanding the leap from reactive Generative AI to goal-driven Agentic AI requires more than just skimming industry updates—it demands a master-level command of data pipelines, statistical modeling, vector memory, and algorithmic design. As artificial intelligence shifts from simple prompt execution to multi-step autonomous workflows, the global tech landscape is rapidly seeking leaders who can architect, deploy, and govern these complex intelligent systems.
The Master of Data Science at Monash University, Indonesia is built on those fundamentals: statistical machine learning, data engineering, and the processing of structured and semi-structured data. Students here don't only study this work – Master of Data Science students presented research at ACL 2025, one of the field's leading international venues, and thesis students work inside active projects on model evaluation, multilingual NLP, interpretable and responsible AI.
If you are ready to move beyond the surface level and engineer the next generation of AI and machine learning solutions, the Master of Data Science at Monash University, Indonesia is your springboard.
Classes run Saturdays at the BSD City campus, with live sessions online on weekday evenings. Applicants from both IT and non-IT backgrounds are welcome – foundation units in Python and SQL bring everyone to the same starting line. Scholarships are available.
This article has been reviewed by
