Autonomous Agents
Autonomous AI agents are software systems that use large language models as a reasoning core, enabling them to plan multi-step tasks, use external tools, maintain memory, and take actions to achieve goals with minimal human intervention.
Autonomous agents — sometimes called agentic AI systems — are AI systems that operate with a degree of self-direction, pursuing goals over multiple steps by planning, using tools, observing outcomes, and adapting their approach without requiring human input at each decision point. Powered by large language models (LLMs) as their reasoning core, these systems represent a qualitative shift from language models that respond to individual prompts toward AI that can execute complex, open-ended workflows. The year 2025 has been widely identified as the inflection point for enterprise agentic AI adoption, with 86% of Malaysian business leaders expressing confidence in using agents to augment workforce capacity within 18 months, significantly exceeding the global average.[^1]
Core Components
An autonomous agent comprises four functional layers that work in concert to bridge the gap between a user goal and its realisation.
Planning
Given a high-level objective, the agent must decompose it into a sequence of actionable sub-tasks. LLMs are well-suited to this role: when prompted to reason step by step (chain-of-thought), they produce structured plans that can be tracked and modified. More sophisticated agents employ reflection loops — after executing each sub-task, the agent evaluates the result against the expected outcome and revises its plan accordingly. This makes agents more robust to unexpected tool responses or partial failures.
Memory
Agents operate across much longer horizons than single-turn chatbots, so memory is a critical architectural concern. Memory in agentic systems is typically divided into: working memory — the contents of the current LLM context window, including the conversation, tool results, and plan state; episodic memory — a retrievable record of past actions and outcomes, often stored in a vector database for semantic retrieval; and semantic memory — structured knowledge such as user preferences, learned facts, and environment state stored in databases or files.
Tool Use
Tools are the interface between an agent and the external world. Through function calling (standardised by the OpenAI API schema and supported natively by models from Anthropic, Google, and Mistral), an agent can invoke APIs, execute code, query databases, read and write files, browse the web, or interact with GUI applications. The agent specifies which tool to call and with what arguments; the result is returned as a tool response and incorporated into the agent's context for subsequent reasoning.
Observation and Feedback
After each action, the agent receives an observation — the output of the tool, an error message, or a state update. The LLM processes this observation to update its understanding of progress toward the goal and determine the next action. This observe–act loop continues until the agent determines the goal has been achieved, reaches a decision point requiring human input, or hits a defined resource or step limit.
Multi-Agent Architectures
Complex tasks can be decomposed across multiple specialised sub-agents that collaborate, with an orchestrator agent routing tasks and synthesising results. Multi-agent frameworks such as CrewAI, AutoGen, and LangGraph support role-based agent design — for example, a research agent, a code generation agent, and a review agent collaborating on a software development task. Communication between agents may be structured as message passing, shared memory, or hierarchical delegation.
Key Frameworks and Platforms
LangChain and LlamaIndex provided the earliest production-grade abstractions for building agents in Python, offering modular components for memory, tools, and agent loops. OpenAI's Assistants API (later the Responses API with native tool use) and Anthropic's Agent SDK enable cloud-hosted agent runtimes. Microsoft Copilot Studio and Salesforce Agentforce offer no-code/low-code platforms for building enterprise agents without custom coding. Amazon Bedrock Agents provides a managed AWS-native agent infrastructure.
Safety and Control
Autonomous operation creates new risk categories. Actions taken by agents — sending emails, modifying databases, calling financial APIs — may be difficult to reverse. Best practices include defining explicit permission boundaries (what tools and resources an agent may access), requiring human-in-the-loop approval for high-stakes actions, implementing step and cost limits, logging all tool calls for auditability, and sandboxing code execution environments. The alignment of agent goals with user intent — ensuring agents do what was meant rather than just what was said — is an active research area.
References
- Microsoft. (2025). 2025 Work Trend Index: The Year the Frontier Firm Is Born. Microsoft Corporation.
- Google & Temasek. (2025). e-Conomy SEA 2025. Google, Temasek, and Bain & Company.
- AIBP. (2026). The Age of Autonomous Finance: Malaysia's Ryt Bank Redefines Banking with Full AI Integration. AI in Business Platform.
- Wang, L., Ma, C., Feng, X., et al. (2024). A survey on large language model based autonomous agents. Frontiers of Computer Science, 18(6).