Agentic AI

Agentic AI describes artificial intelligence systems that operate as goal-directed agents: entities capable of autonomously planning, executing, adapting, and learning across sequences of actions to accomplish complex objectives, without requiring a human prompt at each step. Whereas conventional AI systems operate reactively — receiving a prompt and producing a single response — agentic AI systems decompose high-level goals into subtasks, select and use tools to execute those subtasks, evaluate intermediate results, and revise their plans based on what they observe.^[1] The emergence of agentic AI as a practical paradigm between 2023 and 2025 represents one of the most significant shifts in how AI systems are deployed in enterprise and consumer contexts.

From Reactive to Agentic Systems

The distinction between reactive and agentic AI lies primarily in temporal scope and autonomy. A language model answering "summarise this document" is reactive: it receives input and returns output in a single exchange. An agentic system given the goal "research our three main competitors and produce a comparative pricing analysis" will autonomously conduct web searches, extract data from multiple sources, reconcile conflicting figures, populate a spreadsheet, and deliver a structured report — potentially across dozens of discrete actions spanning several minutes or hours. This shift is made possible by several converging developments: large language models with sufficient reasoning capability to decompose complex goals and self-evaluate progress; tool use and function calling APIs that allow models to interact with external systems including search engines, databases, calendars, and code interpreters; and memory mechanisms that allow context to persist across the action sequence.

The PRAL Loop

Researchers and practitioners often describe agentic AI cognition using the Perceive–Reason–Act–Learn (PRAL) loop.^[2] In the Perceive phase, the agent collects information from its environment: reading files, querying databases, calling APIs, browsing websites, or receiving messages from other agents. In the Reason phase, it analyses the collected information, interprets context, and formulates a plan — deciding which action to take next and in what order. In the Act phase, it executes the chosen action autonomously: writing and running code, sending requests, modifying files, or delegating sub-tasks to other agents. In the Learn phase, the agent incorporates feedback from the outcome of its actions, updating its understanding and adjusting subsequent behaviour. This loop can execute thousands of times in a single agentic run, enabling the system to handle ambiguity, recover from errors, and adapt to unexpected intermediate results in ways that single-shot language model calls cannot.

Core Architectural Components

Planning

Agentic systems require a planning mechanism to decompose goals into executable steps. Common approaches include chain-of-thought reasoning, in which the model writes out its reasoning before acting; tree-of-thoughts search, in which multiple possible action sequences are explored and evaluated; and ReAct (Reason + Act), a prompting framework that interleaves reasoning traces with tool invocations.

Tool Use

Effective agents require access to tools beyond language generation: web search, code execution, file manipulation, calendar and email APIs, form submission, and database query are among the most common. Tool use is enabled through function calling interfaces provided by API providers, which allow the model to emit structured requests for specific tools and receive structured results in return.

Memory

Agent memory spans multiple forms: in-context memory (information within the current context window), external memory (retrieved from vector databases or other storage systems), episodic memory (records of past interactions and outcomes), and semantic memory (distilled facts about the world or user preferences). Long-running agents require robust memory systems to maintain coherent task state across time.

Multi-Agent Orchestration

Complex tasks are often distributed across multiple specialised agents coordinated by an orchestrator. An orchestrator agent might delegate research to a browsing agent, code generation to a programmer agent, and data formatting to a document agent, synthesising their outputs into a final deliverable. Frameworks such as LangChain, LlamaIndex, AutoGen, and CrewAI provide scaffolding for building these multi-agent architectures.

Applications

Agentic AI has found applications across a wide range of enterprise domains. In software engineering, coding agents can receive a feature specification, write implementation code, run tests, interpret failures, and iterate until tests pass — substantially reducing the manual effort of routine development tasks. In business operations, agents can process invoices, reconcile accounts, schedule meetings, respond to customer queries, and escalate edge cases to human staff. In research and analysis, agents can survey literature, extract data from reports, and produce structured summaries. Gartner estimated in 2025 that one-third of enterprise software would embed agentic AI capabilities by 2028.^[3]

Safety and Control Challenges

The autonomy that makes agentic AI valuable also introduces novel safety challenges. An agent executing dozens of actions across external systems can cause cascading errors that are difficult to reverse — deleting data, sending unintended communications, or incurring unexpected API costs. Containment strategies include sandboxed execution environments, confirmation checkpoints for irreversible actions, rate limits on external API calls, and human-in-the-loop review for high-stakes decisions. The field of agentic AI safety is an active research area, with particular concern around agents that misinterpret goals (specification gaming), take unintended shortcuts, or act in ways that circumvent intended oversight. Malaysia's digital transformation agenda, articulated through the MyDigital Blueprint and the National AI Roadmap, explicitly targets the deployment of intelligent automation across government services, financial services, and key economic sectors including manufacturing and logistics. Agentic AI systems represent the leading edge of this automation capability, moving beyond simple chatbots and rule-based workflows toward systems capable of end-to-end task execution. Malaysian banks including Maybank and CIMB have piloted agentic AI systems for back-office automation, including loan document processing, fraud investigation workflows, and regulatory reporting. BNM's guidance on responsible AI in financial services emphasises the importance of human oversight and audit trails for automated decision processes — requirements that agentic AI deployments must satisfy through logging, confirmation checkpoints, and explainability mechanisms. Grab Malaysia and AirAsia have deployed agent-based systems in logistics and customer operations: coordinating driver assignment, managing flight disruption communications, and handling high-volume customer service interactions that require retrieving booking data, checking policy eligibility, and executing refund workflows. These use cases illustrate both the efficiency gains and the safety requirements associated with agents that can take irreversible actions on behalf of customers. The Malaysian government's MAMPU (Malaysian Administrative Modernisation and Management Planning Unit) digital government programme has evaluated agentic AI for public service automation, including permit application processing and benefits eligibility determination. Pilot deployments are subject to the AI Governance Framework's requirements for human oversight of consequential decisions affecting citizens. MDEC has also incorporated agentic AI literacy into its Digital Leadership Academy, recognising that managers need to understand agent capabilities and limitations to deploy them responsibly.

1. Wrike. (2025). What is Agentic AI? The Autonomous Future of Work Explained. Wrike Blog.
2. KPMG India. (2025). Agentic AI: The Future of Autonomous Intelligence. KPMG Insights.
3. Gartner. (2025). Gartner Predicts One-Third of Enterprise Software Will Embed Agentic AI by 2028. Gartner Press Release.
4. Yao, S., et al. (2022). ReAct: Synergizing Reasoning and Acting in Language Models. arXiv:2210.03629.