Indonesian
Arabic
Chinese (Simplified)
Dutch
English
French
German
Italian
Portuguese
Russian
Spanish
Autonomous generative AI agents execute complex tasks with little or no human supervision. Agentic AI differs from chatbots and co-pilots.
Unlike traditional AI, particularly generative AI, which often requires human intervention in complex workflows, agentic AI aims to autonomously navigate and optimize processes thanks to its decision-making capabilities and goal-directed behavior. AI agents serve as:
AI code editors like Cursor AI Editor, Windsurf Editor, andReplit aims to build and deploy apps (e.g. To-Do list app) by:
A developer used OpenAI’s Operator and Replit’s AI Agent to build an entire app in 90 minutes. Two agents autonomously exchanged credentials, and ran tests.
Cursor’s agent mode Composer aims to generate a complete Tic Tac Toe game from a single prompt:“Generate an HTML, CSS, and JavaScript Tic Tac Toe game for 2 players.”
Cursor is cable of coding across multiple files, execute commands, and automatically determine what context it needs (no need to add files).
If you are exploring the infrastructure powering agentic AI systems, we recommend checking out our latest benchmarks:
AI code editors automate API creation by transforming specifications into functional code. Here’s how the process typically works:
No-Code API workflows for AI Agents with n8n.
Here is a high-level abstraction of an API workflow:
You can select code snippets and issue plain English commands such as:
“Double the size of the board. Make it green – like an Apple 2e.” (see real-life example below).
Coding agents like Cursor identify the intent, modifies the relevant code across files, and applies the changes.
AI website creators like v0 by Vercel, Bolt, Lovable, and CerebrasCoder aims to create complex platforms like e-learning websites, generating key pages such as:
Similarly, AI code editors like Replit builds websites and leverage APIs.4
How to build a website AI agent:
Here is the automation workflow:
Agents are capable of generating the front-end interface, configure back-end logic, and set up database interactions.
Roo Code uses DeepSeek model to autonomously build complete CRM dashboards.
While basic coding is typically a task for LLMs, recursive coding workflows where an agent iteratively improves/extends code across multiple layers are inherently agentic.
Agents autonomously rewrite large code blocks, apply configuration changes, and test outcomes in cycles until a goal is met.
GT Edge AI converts legacy COBOL code into modern Java.7
Persistent provides a multi-agent framework used to autonomously migrate COBOL code to Java, it works by:8
By using recursive coding AI agents continuously improve the design of the code without changing how it works, making it easier to understand and maintain.
Tech startup developers created an agent refactoring your code in 25+ programming languages.9
Agents like GitHub Copilot provide real-time code suggestions and auto-completions and reduce the likelihood of syntax errors.
AI agents manage infrastructure in cloud-native environments like Kubernetes. These DevOps agents aims to:
When connected to Kubernetes via tools or wrappers, Claude can act as a DevOps agent for querying cluster state. “Check if I have any pods running.”.
Agents gather and correlates threat actor TTPs (tactics, techniques, procedures) from open-source and proprietary feeds, and integrates findings into detection workflows.
Microsoft’s Security Copilot includes a specialized Threat Intelligence Briefing Agent that dynamically gathers, filters, and summarizes threat intelligence.
These actions occur at the initial signal ingestion stage to reduce noise and organize alert data before deeper analysis.
Charlotte AI performs autonomous detection and triage by:
Performed after initial triage, this step adds depth and context to alerts.
Automated attacker attribution systems ingest CTI feeds, extract behavioral and temporal features, and compute similarity scores across incidents. Clustering algorithms then map intrusions to known threat actors like (e.g., APT41, Mozi or Lazarus) based on pattern overlap.
Google Chronicle + Mandiant + Gemini AI agents autonomously ingest telemetry and CTI feeds, enrich alerts with IOC context (e.g., IP reputation, malware hashes), and cross-reference behavioral patterns with known threat actor tactics from the MITRE ATT&CK framework.
In this agentic setup:
In SecOps, agents isolate endpoints, disable accounts, or kill malicious processes., these systems aims to:
Google developed the SOC Manager agent, which leverages multiple sub-agents to execute a structured Incident Response Plan for malware detection.16
Agentic project structure:
In the final step of the incident response plan (Step 5: Completion), the IOCs (Indicators of Compromise) were proactively blocked by the SOC Manager agent executing an automated containment runbook (see below).
Source:
Mandiant & Google Cloud Security17
Agents in threat hunting aims to:
Researchers developed a MITRE ATT&CK Driven Threat Hunting Automated by Local LLM system, where AI agents collaborate to generate Sigma rules for threat detection.
In this example A user inputs a request (e.g., “Please generate a Sigma rule for hunting Kerberoasting”) through a web UI.
Agent 1 retrieves relevant detection methods from MITRE ATT&CK, while Agent 2 uses this input to generate context-aware Sigma rules using a language model.18
AI testing agents create and execute unit, integration, vulnerability, and performance tests without extensive manual intervention. However, building these AI systems is resource-intensive, as they require significant computational power.
Pcloudy’s Copilot provides selenium test scripts, and finds available browsers to test on and execute the test cases.19
Read more: Enterprise AI assistants, AI agent builders, open-source AI agents.
AI agents improve NPCs and other agentic processes in the game world by performing NPC behaviors, game playing & adaptability, and procedural content generation.
Fully autonomous AI agents in gaming provide human-like behavior and gameplay for non-player characters (NPCs)..
Researchers created a small virtual town populated with AI by building a sandbox setting similar to The Sims with 25 agents called “Stanford AI Village”.
In this village, users can observe and interact with agents as they share news, build relationships, and arrange group activities. 20
Here’s an overview of the key components and ideas behind these concepts:
AI agents play video games or assist human players in achieving specific goals by leveraging:
Google DeepMind’s Scalable Instructable Multi-Agent (SIMA) navigate and interact with gaming situations. SIMA aims to play games such as No Man’s Sky and Goat Simulator.
Source: Google22
AI agents are highly capable of generating vast amounts of game content algorithmically, such as:
No Man’s Sky, an adventure game, uses procedural generation to create entire planets with:
AI agents automate content creation, editing, and publishing. These AI agents assist human writers, and generate content independently. Some applications of AI writing assistants include:
AI agents write a narrative, by outlining chapters, drafting content, and polishing prose chapters, drafting content, and polishing prose.
In a GitHub AI agent project, 10 specialized AI agents worked autonomously to write a novel of 100,000 words (~300 pages) with zero human writing. Some examples from those 10 agents include:
Here is a Livestream showing how agents create the novel:
AI agents autonomously draft technical reports including:
ParagraphAI, an AI writing assistant, write technical engineering reports by outlining the timeline, the budget, and the resources and personnel required.25
Agents pull information from knowledge databases like Wikipedia, product manuals, or academic journals to create a comprehensive overview of a specific topic.
Perplexity Pages turns gen AI search results into structured Wikipedia pages.26
Agents generate UI/UX components, system diagrams, and flowcharts based on text prompts, streamlining the design process.
FigJam AI uses text prompts to generate:
Agents automate claims review, approval, and fraud detection, streamlining the entire claims processing lifecycle. For example, a large-scale insurer automates ~90% of individual automobile claims by integrating custom AI agents into their claims workflow. 28
Once a claim is submitted, an agentic AI systems extract relevant data from the submitted forms, verify the details against existing databases, and flag any inconsistencies or potential fraud signals.
Microsoft’s Power Platform – automates an insurance claim form:
Here, Microsoft’s Power Platform:
Agentic AI automate underwriting with specialized agents, including a risk evaluation agent for claim likelihood assessment, and a pricing agent for dynamic premium adjustments, etc.
Akira AI‘s agents automate insurance underwriting and risk assessment through the multi-agent system, each specializing in a critical aspect of underwriting:
Agents communicate the claim status and next steps to the policyholder, including any additional documentation required, approval/denial updates, or payout details.
HR operations often involve numerous repetitive like resume screening tasks that can be automated. Here are key examples of agentic AI in HR operations:
Agentic workflows automate the screening process, filter relevant skills, and automatically assign scores based on your pre-defined criteria.
PepsiCo uses AI tools to rank candidates depending on how well they meet job requirements.31
Agents can handle the scheduling of interviews, and coordinating between candidates and hiring managers to find optimal times.
LinkedIn HR Assistant performs day-to-day tasks like synthesizing job descriptions, searching for candidates, and doing basic screening calls.32
Agents in payroll processing calculate salaries, process deductions, and handle tax withholdings. They integrate with human resources information system (HRIS) systems and accounting software to ensure accuracy and compliance with payroll standards.
Explore more on finance automation solutions like:
Akira AI’s multi-agent payroll system automates every aspect of the payroll cycle. The system uses several agents, including:
Source: Akira AI33
Here, Akira AI’s multi-agent payroll system uses several agents, including:
Traditional chatbots answer basic questions, but they often hit a wall when it comes to actually helping the customer. Agentic customer service tools changes that by:
When a customer calls about an inquiry, AI agents process the call with natural language.
Ada AI Agent answers customer calls:
AI agents deliver context-specific responses or direct customers to the appropriate resources for further assistance.
After an interaction, agents send SMS messages to follow up with customers.
Agents:
AI agents as research assistants are used in various fields to assist with data analysis, literature review, hypothesis generation, and experimental design.
1. OpenAI’s Deep Research uses reasoning to synthesize large amounts of online information and complete multi-step research at a Ph.D level when running large searches using the o3 & DeepSeek.35 36
In an experiment where researchers asked Deep Research to conduct a real-life project with a detailed prompt, Deep Research:
2. ChemicalQDevice’s clinical decision support (CDS) system was asked to execute agentic workflow for drug discovery. In this example ChemicalQDevice’s system:
3. End-to-end agentic workflow system, otto-SR, leverage LLMs to conduct literature searches, applies inclusion/exclusion criteria, extracts structured data, and performs meta-analyses.39
OpenAI’s Deep Research ChemicalQDevice’s system and otto-SR can be used in several agentic use cases given below:
Autonomously searching academic databases, journals, and online research repositories (e.g., Google Scholar, PubMed) to gather relevant studies, papers, and articles related to specific research topics or hypotheses.
Proactively generating analysis hypotheses based on patterns in the data and testing them (work that analysts and business users typically do).
Manipulating structured and unstructured data from several sources like research databases, social media, patents, or clinical trial results, providing insights into emerging trends.
Generating insightful visual representations of complex datasets.
“Computer Use” aims to enable AI to interact with a computer like a person would. This gives the flexibility to perform digital tasks without using OS- or web-specific APIs.
There are two approaches for AI agents to perform tasks like humans:
Tool examples:
Agents navigate webpages, click fields, and fill out forms based on user prompts or structured data.
Agents open files, make edits, rename, organize, and save documents across local or cloud environments.
Anthropic’s Claude is asked to “Generate 25 rows of sample expenses, save them into a spreadsheet, and then open the spreadsheet”.
In this example, Claude:
Unlike basic automation scripts, deep web research agents interpret unstructured information across multiple pages and return insights in a structured format.
OpenAI’s Deep Research, a new agentic capability within ChatGPT designed for multi-step, high-context web research plans, navigates, and synthesizes information across multiple sources to answer complex queries.
CLI-based coding agents like Aider that are designed for terminal-based development workflows run shell commands, install software, launch scripts, and interpret outputs in terminal interfaces.
Aider, a CLI-native AI coding agent, is used by developers to refactor codebases and execute shell commands such as running test suites (pytest, npm test). The agent interprets terminal outputs, fixes errors iteratively, and commits changes directly to Git repositories.48
Unified GUI agents (e.g.,OpenAI Operator prototypes) can switch between applications.
To test an order delivery use case, I provided a simple shopping request from Open Operator: Help me buy a boho-style throw pillow cover under $30.
Planning agents with memory + tool use (e.g., Auto-GPT, Agent Q with goal-setting) execute multi-step goals across varied tools (e.g., gather input, take actions, revise plan), making decisions in real time.
In thus multi-step financial report analysis the AI planning agent is given the task: “Analyze last quarter’s financial performance and prepare a summary for the finance team.”
The agent is asked to get:
Here is the financial report output:
MultiOn Agent Q booking a flight.
The shift from basic agent applications, such as natural language weather searches using tools like LangChain, to more complex, autonomous use cases like app development (e.g., generating a Tic Tac Toe game with the Cursor AI editor) resulted in the challenges:
AI agent-building frameworks help address these challenges by providing:
Creating a custom Slovenia trip guide agent with Microsoft 365:
While enterprises are running many PoCs on the topic, cost of mistakes are quite high in numerous enterprise workflows. The probabilistic nature of LLMs reduces their reliability and slows down the adoption of agents in production environments.
Agentic AI is the overarching framework that enables AI systems to solve problems with minimal supervision. Within this framework, AI agents are the individual components responsible for executing specific tasks autonomously.
While agentic AI understands user goals and orchestrates the problem-solving process, AI agents perform the tasks.
Decision-Making: Operates with minimal human input by assessing situations and choosing actions based on predefined goals and evolving context.
Problem-solving: Follows a four-step loop: perceive → reason → act → learn.
Autonomy: Agentic AI systems act independently, learning and improving over time.
Interactivity: Proactively engages with its environment, adjusting actions in real-time (e.g., self-driving cars making dynamic driving decisions).
Planning: Capable of executing multi-step strategies, allowing it to solve complex tasks and achieve long-term objectives.
Generative AI creates content on request, while agentic AI pursues goals independently.
Generative AI creates original content based on user prompts. It is reactive, responding to specific user input. Tools like ChatGPT and GitHub Copilot are popular examples.
Agentic AI, in contrast, is designed to act autonomously. It combines LLMs with tools like reinforcement learning and knowledge representation to make decisions, plan steps, and adapt to changing situations. It’s proactive, capable of initiating and completing complex tasks.
Your email address will not be published. All fields are required.