AI agents
Last reviewed
May 8, 2026
Sources
26 citations
Review status
Source-backed
Revision
v8 · 7,850 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 8, 2026
Sources
26 citations
Review status
Source-backed
Revision
v8 · 7,850 words
Add missing citations, update stale details, or suggest a clearer explanation.
AI agents are software systems that use artificial intelligence to perceive their environment, reason about goals, make decisions, and take autonomous actions to accomplish tasks on behalf of a user or another system. Unlike traditional AI models that simply generate responses to prompts, agents operate in loops of perception, reasoning, and action, often using external tools and maintaining memory across interactions. The term has become central to the AI industry since 2023, when large language models (LLMs) enabled a new generation of agents capable of general-purpose reasoning, planning, and tool use.
The concept of an agent has been a foundational abstraction in AI since the field's earliest days. In the most widely cited textbook definition, Stuart Russell and Peter Norvig describe an agent as "anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators." Their textbook, Artificial Intelligence: A Modern Approach (1995, fourth edition 2021), frames the entire study of AI as "the study and design of rational agents," where a rational agent "acts so as to achieve the best outcome or, when there is uncertainty, the best expected outcome."
In practice, AI agents range from simple thermostat-like systems that follow fixed rules to sophisticated LLM-powered autonomous systems that can browse the web, write and execute code, manage files, and interact with external APIs. As of 2026, AI agents represent one of the fastest-growing segments of the AI industry. The AI agents market grew from $5.4 billion in 2024 to roughly $7.6 billion in 2025, with projections reaching $50 billion by 2030. Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025.
There is no single agreed-upon definition of an AI agent, but most researchers and practitioners identify several characteristics that distinguish agents from simpler AI systems.
An AI agent is a system that can:
Michael Wooldridge and Nicholas Jennings, in influential work from the 1990s, identified four properties that distinguish intelligent agents: autonomy (operating without direct human control), reactivity (responding to changes in the environment), proactiveness (taking initiative to achieve goals), and social ability (interacting with other agents or humans).
The key distinction between an agent and a standard chatbot or LLM is the action loop. A chatbot receives a prompt and returns a response. An agent receives a goal, then enters a loop where it repeatedly reasons about what to do next, takes an action, observes the result, and decides whether to continue or stop. This loop can run for seconds, minutes, or even hours depending on the complexity of the task.
Andrew Ng, a prominent AI researcher, has described the spectrum of "agentic" behavior as a continuum rather than a binary classification. A system that uses an LLM to generate a single response is not agentic. A system that uses an LLM to generate a response, then reflects on that response, then revises it, is somewhat agentic. A system that autonomously plans multi-step workflows, uses tools, and adapts based on results is highly agentic.
The concept of software agents predates modern AI by decades. The earliest agent-like systems emerged during the Symbolic AI era, when researchers built programs that could make autonomous decisions within narrow domains. Alan Turing's 1950 paper "Computing Machinery and Intelligence" proposed the Turing test as a way to evaluate machine intelligence, and the 1956 Dartmouth Conference, organized by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, established AI as a research discipline.
ELIZA, developed by Joseph Weizenbaum at MIT in 1966, was one of the first programs to simulate conversational behavior, using pattern matching to mimic a psychotherapist. SHRDLU (1970), built by Terry Winograd at MIT, could understand natural language commands and manipulate objects in a simulated "blocks world," acting as an agent within a microworld. STRIPS (Stanford Research Institute Problem Solver, 1971) introduced formal planning capabilities, allowing an agent to reason about actions and their effects to achieve goals.
The expert systems of the 1970s and 1980s represented a more sophisticated form of agent behavior. DENDRAL (1965 to 1983) proposed molecular structures for organic compounds based on mass spectrometry data, while MYCIN (1972 to 1980) diagnosed infectious diseases and recommended treatments using approximately 600 hand-crafted rules. These systems could reason within their domains, but they were brittle, expensive to build, and unable to generalize beyond their programmed knowledge.
Reinforcement learning (RL) introduced a formal framework for agents that learn from interaction with an environment. In RL, an agent takes actions in an environment, receives rewards or penalties, and adjusts its behavior to maximize cumulative reward. This established the classic agent loop (perceive, act, receive feedback, update policy) that would later inspire modern agentic AI workflows.
Craig Reynolds developed "boids" in 1986, demonstrating that simple individual agent rules could produce complex emergent group behavior, an early example of multi-agent systems. Academic research on multi-agent systems expanded through the 1990s and 2000s, exploring coordination, negotiation, and competition between autonomous software entities.
DeepMind's work on game-playing agents marked a turning point. IBM's Deep Blue defeated chess world champion Garry Kasparov in 1997, and AlphaGo defeated world champion Lee Sedol at Go in 2016. Later systems like AlphaStar (StarCraft II) and OpenAI Five (Dota 2) demonstrated that RL agents could master complex, multi-step strategic environments.
The 2010s also saw the emergence of consumer virtual assistants such as Apple's Siri (2011), Amazon's Alexa (2014), and Google Assistant (2016). While these systems could carry out simple voice-driven tasks, they relied heavily on predefined intents and lacked the ability to reason through multi-step problems or use arbitrary tools.
The release of GPT-3 in 2020 and ChatGPT in late 2022 revealed that large language models could serve as general-purpose reasoning engines. Early experiments with prompt engineering showed that LLMs could simulate procedural reasoning when given instructions like "think step by step."
In October 2022, researchers at Princeton and Google published the ReAct paper ("Synergizing Reasoning and Acting in Language Models"), which introduced a framework where an LLM alternates between generating reasoning traces and taking actions. ReAct demonstrated that interleaving thought and action dramatically improved performance on knowledge-intensive tasks and interactive environments compared to reasoning or acting alone.
In February 2023, Meta AI published the Toolformer paper (Schick et al.), showing that language models could teach themselves to use external tools such as search engines, calculators, and translation systems via API calls. Toolformer established that tool use did not need to be hand-programmed; models could learn when and how to invoke tools through self-supervised training, with API calls inserted into the training data at useful positions.
The spring of 2023 saw an explosion of open-source autonomous agent projects. AutoGPT, created by Toran Bruce Richards and released in March 2023, became one of the fastest-growing GitHub repositories in history, amassing over 100,000 stars within months. AutoGPT worked by taking a user-defined goal, breaking it into sub-tasks, and using GPT-4 to execute them autonomously, with access to web search, file operations, and code execution.
BabyAGI, a Python script created by venture capitalist Yohei Nakajima, went viral around the same time. It orchestrated a simple loop of task creation, execution, and prioritization using an LLM and a vector memory store. Despite their limitations (frequent loops, high API costs, unreliable outputs), AutoGPT and BabyAGI proved the concept of LLM-powered autonomous agents and sparked massive investment and research in the field.
By 2024, the AI industry had shifted decisively toward building production-grade agent systems. Major labs released dedicated agent APIs, frameworks matured, and enterprises began deploying agents for real workflows. OpenAI launched its Assistants API (later succeeded by the Responses API and Agents SDK), Anthropic introduced Claude with computer use capabilities, and Google released Project Mariner for autonomous web browsing.
The year 2025 marked the emergence of standardized protocols (MCP, A2A) and governance structures (the Agentic AI Foundation) designed to make agents interoperable and safe at scale. By early 2026, agents had moved from experimental demos to production deployments at thousands of enterprises.
AI agents can be classified along several dimensions. The most widely referenced taxonomy, drawn from Russell and Norvig's textbook, identifies five types based on increasing sophistication.
Simple reflex agents select actions based solely on the current percept, ignoring the rest of the percept history. They operate on condition-action rules (also called "if-then" rules): if the current input matches a condition, the agent performs the corresponding action. A thermostat is a classic example. These agents work well only in fully observable environments where the correct action can be determined from the current state alone.
Model-based reflex agents maintain an internal model of the world that tracks aspects of the environment not directly observable. This internal state, combined with the current percept, allows the agent to handle partially observable environments. For example, a self-driving car maintains a model of nearby vehicles even when they temporarily leave the sensor range.
Goal-based agents extend model-based agents by incorporating explicit goal information. Rather than simply reacting to the current state, they consider future states and evaluate which sequences of actions will achieve their goals. This requires search and planning capabilities. A navigation system that computes a route to a destination is a goal-based agent.
Utility-based agents go beyond simple goal achievement by assigning a utility value (a measure of "happiness" or desirability) to different states. When there are multiple ways to achieve a goal or when goals conflict, the utility function allows the agent to choose the action that maximizes expected utility. A stock trading agent that balances risk against returns is operating as a utility-based agent.
Learning agents can improve their performance over time through experience. They consist of four conceptual components: a performance element (which selects actions), a learning element (which modifies the performance element based on feedback), a critic (which provides feedback based on a fixed performance standard), and a problem generator (which suggests exploratory actions). Most modern AI systems, including LLM-based agents, incorporate learning in some form.
| Agent type | Internal state | Planning | Learning | Example |
|---|---|---|---|---|
| Simple reflex | None | No | No | Thermostat |
| Model-based reflex | World model | No | No | Spam filter with context state |
| Goal-based | World model + goals | Yes | No | Route planner |
| Utility-based | World model + utility function | Yes | No | Financial trading agent |
| Learning | All of the above + learning element | Yes | Yes | Self-driving car, AlphaGo |
AI agent architectures define how an agent processes inputs, maintains state, makes decisions, and executes actions. Several design patterns have emerged as the field has matured.
A typical LLM-based agent operates in a loop:
This perceive-think-act-observe loop can run for anywhere from a single iteration (simple question answering) to hundreds of iterations (complex software engineering tasks that involve reading code, writing patches, running tests, and debugging failures).
The ReAct pattern, introduced by Yao et al. in 2022, interleaves reasoning steps with action steps. At each iteration, the agent generates a "thought" explaining its reasoning, then selects and executes an action (such as a tool call), then observes the result. This cycle repeats until the task is complete.
ReAct's key advantage is interpretability. Because the agent explicitly articulates its reasoning before each action, humans can follow the agent's decision-making process and identify errors. ReAct has become the default pattern for many agent frameworks, including early versions of LangChain agents.
In the plan-and-execute pattern, the agent first generates a complete plan for accomplishing a goal, then executes each step sequentially. If a step fails or produces unexpected results, the agent may re-plan from that point.
This pattern works well for tasks with clear structure and predictable steps, such as data analysis pipelines or multi-step research workflows. The separation of planning from execution also makes it easier to audit and control agent behavior. LangGraph's "plan-and-execute" template is a widely used implementation of this pattern.
Reflection agents add a self-critique step after each action or after completing a draft output. The agent generates output, then evaluates that output against criteria (correctness, completeness, style), then revises based on its own feedback. This loop can repeat multiple times.
Reflection significantly improves output quality for tasks like writing, code generation, and analysis. Anthropic's Claude models, for example, use extended thinking (also called "thinking" or "scratch pad") to reason through complex problems before producing a final answer. Reflexion (Shinn et al., NeurIPS 2023) added a verbal self-critique step where the agent reviews its own outputs, writes a reflection to memory, and uses that reflection in subsequent attempts.
Tool use enables agents to interact with external systems, moving beyond text generation to real-world action. Common tools include web search APIs, code interpreters, database queries, file system access, browser automation, and custom API integrations.
The mechanism works through function calling (also called tool calling). The developer provides the agent with a set of available tools, each described by a name, a natural language description, and a JSON schema specifying its parameters. The model then decides whether to call a tool and, if so, which tool to call with what arguments. The runtime executes the call and returns the result, which is fed back into the model's context for the next decision. Modern LLMs are trained with function-calling capabilities that make tool use reliable and structured. OpenAI's function calling, Anthropic's tool use API, and Google's function calling for Gemini all provide standardized interfaces for this pattern.
Rather than building a single agent that handles everything, multi-agent systems use specialized agents that collaborate on complex tasks. Common orchestration patterns include:
| Pattern | Description | Best for |
|---|---|---|
| Hierarchical | A supervisor agent delegates tasks to specialized worker agents and aggregates their results | Complex workflows with clear task decomposition |
| Hub-and-spoke | A central orchestrator routes messages between agents, maintaining a shared state | Workflows requiring strong consistency and auditability |
| Mesh / peer-to-peer | Agents communicate directly with each other without a central coordinator | Resilient systems that need to handle partial failures gracefully |
| Pipeline | Agents process work sequentially, each transforming the output of the previous agent | Linear workflows like content creation (research, write, edit, review) |
| Debate / adversarial | Multiple agents propose solutions and critique each other's outputs | Tasks requiring high accuracy where errors are costly |
Gartner reported a 1,445% surge in multi-agent system inquiries from Q1 2024 to Q2 2025, reflecting growing enterprise interest. Organizations using multi-agent architectures report 45% faster problem resolution and 60% more accurate outcomes compared to single-agent systems.
A growing ecosystem of frameworks, SDKs, and platforms supports agent development. The following table summarizes the major options as of early 2026.
| Framework / Platform | Organization | First Release | Key Features | License |
|---|---|---|---|---|
| LangChain / LangGraph | LangChain Inc. | October 2022 | Graph-based workflows, stateful agents, 47M+ PyPI downloads, largest ecosystem of integrations | MIT |
| AutoGPT | Significant Gravitas | March 2023 | Autonomous goal-driven agents, web browsing, file operations, 160k+ GitHub stars | MIT |
| CrewAI | CrewAI Inc. | December 2023 | Role-based multi-agent teams, natural task division, low barrier to entry | MIT |
| AutoGen / AG2 | Microsoft Research | September 2023 | Multi-agent conversation patterns, group chat, nested conversations (maintenance mode as of 2025) | MIT |
| Microsoft Semantic Kernel | Microsoft | March 2023 | Enterprise-grade LLM integration with C# and Python, planners, plugins | MIT |
| OpenAI Agents SDK | OpenAI | March 2025 | Tool use, handoffs, guardrails, tracing; production successor to Swarm; provider-agnostic | MIT |
| OpenAI Responses API | OpenAI | March 2025 | Unified API merging Chat Completions and Assistants capabilities; built-in web search, file search, computer use | Proprietary |
| Claude Agent SDK | Anthropic | 2025 | Same infrastructure powering Claude Code; supports building custom agents | Proprietary |
| Google Agent Development Kit (ADK) | 2025 | Multi-agent orchestration, integration with Gemini models and Google Cloud | Apache 2.0 | |
| Amazon Bedrock Agents | Amazon | 2023 | Managed agent service with knowledge bases, action groups, and guardrails on AWS | Proprietary |
| Llama Index | LlamaIndex Inc. | November 2022 | Data indexing, retrieval-augmented generation, agent workflows | MIT |
LangChain, created by Harrison Chase, launched in October 2022 and quickly became the most popular framework for building LLM-powered applications. It provides abstractions for chains (sequential LLM calls), agents (LLM + tools), memory, and retrieval. LangGraph, released as a companion library, represents agent workflows as directed graphs where nodes are computation steps and edges define control flow.
By late 2025, LangGraph reached version 1.0 and became the default runtime for all LangChain agents. The framework has accumulated over 47 million PyPI downloads and the largest third-party integration ecosystem in the agentic AI space.
CrewAI, created by Joao Moura and first released in December 2023, focuses on multi-agent collaboration using a role-based metaphor. Developers define "crews" of agents, each with a specific role (researcher, writer, reviewer), and CrewAI orchestrates their interaction. The framework has the lowest barrier to entry for multi-agent prototyping and has grown rapidly through 2025.
AutoGen, developed by Microsoft Research, pioneered structured multi-agent conversation patterns where agents interact through two-agent chats, group chats, sequential conversations, and nested dialogues. In 2025, Microsoft shifted AutoGen to maintenance mode and launched the broader Microsoft Agent Framework, which encompasses Semantic Kernel, Azure AI Agent Service, and Copilot Studio. General availability is targeted for Q1 2026, offering production SLAs and multi-language support.
In March 2025, OpenAI released the building blocks of its new agents platform: the Responses API (merging capabilities from Chat Completions and Assistants APIs), built-in tools for web search, file search, and computer use, and the open-source Agents SDK with tracing. The Assistants API is scheduled for deprecation on August 26, 2026.
The Agents SDK is a production-ready upgrade of OpenAI's earlier experimental framework, Swarm. It provides building blocks for tool use, handoffs between agents, guardrails, and observability, and is designed to be provider-agnostic.
The Model Context Protocol (MCP) is an open standard announced by Anthropic in November 2024 for connecting AI models to external data sources and tools. MCP defines a client-server architecture where an AI application (the MCP client) communicates with tool providers (MCP servers) through a standardized JSON-RPC 2.0 interface, reusing architectural ideas from the Language Server Protocol (LSP).
MCP solves the "N times M" integration problem. Without a standard protocol, every AI application must build custom integrations with every tool. With MCP, tool providers implement a single server, and AI applications implement a single client, and they all work together.
Adoption of MCP has been rapid. OpenAI integrated MCP across its products, including the ChatGPT desktop app, in March 2025. At Microsoft Build 2025, GitHub and Microsoft announced they were joining MCP's steering committee, and Microsoft announced MCP support in Windows 11. Google DeepMind confirmed support in April 2025. Monthly SDK downloads grew from approximately 2 million at launch to over 97 million by late 2025, a roughly 4,750% increase in 16 months. By February 2026, the official MCP registry listed over 6,400 MCP servers.
In December 2025, Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation, establishing open governance for the protocol's future development.
The Agent2Agent Protocol (A2A) was introduced by Google in April 2025 as a communication standard for multi-agent systems. While MCP standardizes how agents connect to tools, A2A standardizes how agents communicate with each other.
A2A uses JSON-RPC 2.0 over HTTPS and defines mechanisms for agent discovery through "Agent Cards" (JSON documents that describe an agent's capabilities, authentication requirements, and connection details), task management, secure information exchange, and coordination across different frameworks and vendors. Version 0.3, released in July 2025, added gRPC support, signed security cards, and an extended Python SDK. The protocol received support from over 150 organizations, including Atlassian, Salesforce, SAP, ServiceNow, and PayPal.
However, adoption of A2A has been slower than MCP. As of late 2025, most of the AI agent ecosystem consolidated around MCP for tool connectivity, and A2A's development pace slowed. The Linux Foundation launched A2A as an open-source project under the Apache 2.0 license to encourage broader participation.
The Agentic AI Foundation (AAIF) was announced on December 9, 2025, as a directed fund under the Linux Foundation. It was co-founded by Anthropic, Block, and OpenAI, with the goal of providing neutral, open governance for the standards and tools that power agentic AI.
The three founding projects are:
Platinum members of the AAIF include Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI. Together, MCP and A2A are forming the foundation of an interoperable agent ecosystem, sometimes compared to how HTTP and SMTP standardized web and email communication.
OpenAI has progressively added agentic capabilities to ChatGPT. The product now includes browsing (powered by Bing), code execution via a Python sandbox, image generation with DALL-E, and file analysis. In late July 2025, OpenAI introduced a Tools dropdown providing access to six specialized modes: Agent mode, Deep research, Create image, Study and learn, Web search, and Canvas.
Deep Research, launched in February 2025, operates as an autonomous research agent that browses the web for 5 to 30 minutes, synthesizing findings into structured reports with citations. Operator, launched January 23, 2025, is powered by a Computer-Using Agent (CUA) model that combines GPT-4o vision with reinforcement learning on GUI control; at launch it scored 38.1% on OSWorld, 58.1% on WebArena, and 87% on WebVoyager.
Anthropic introduced computer use capabilities for Claude in October 2024, allowing the model to interact with desktop environments through screenshot analysis and mouse/keyboard control. The feature enables Claude to operate software, navigate websites, fill out forms, and perform multi-step desktop tasks. Anthropic explicitly framed the launch as experimental, noting it was "at times cumbersome and error prone." Asana, Canva, Cognition, DoorDash, Replit, and The Browser Company were the first listed adopters.
On the OSWorld benchmark for real-world computer tasks, Claude Sonnet 4.5 leads at 61.4%. Claude Code, launched in May 2025, is Anthropic's terminal-based agentic coding tool built on the Claude Agent SDK. By early 2026, Claude Code had achieved a 46% "most loved" rating among developers. In January 2026, Anthropic released Claude Cowork, a graphical agent interface aimed at non-technical users.
Devin, created by Cognition Labs and announced in March 2024, was described as the "first AI software engineer." Devin operates as an autonomous coding agent with its own code editor, browser, and terminal, capable of planning and executing multi-step software engineering tasks. At launch it resolved 13.86% of issues on the full SWE-bench benchmark, far above the previous best unassisted result of 1.96%.
Devin's annual recurring revenue grew from approximately $1 million in September 2024 to roughly $73 million by June 2025. In July 2025, Cognition acquired Windsurf (formerly Codeium), an AI-powered IDE, bringing combined ARR to approximately $150 million. Devin is used by engineering teams at thousands of companies, including Goldman Sachs, Santander, and Nubank.
Manus, launched by the Singapore-based startup originally founded as part of Chinese company Butterfly Effect (Monica.im) in March 2025, is a general-purpose agent built on a multi-agent architecture running in dedicated virtual machines. It scored 86.5%, 70.1%, and 57.7% on GAIA Levels 1, 2, and 3 respectively, exceeding OpenAI's Deep Research scores reported at the same time. Within a week of launch, more than two million people joined its waitlist. In late December 2025, Meta announced the acquisition of Manus in a deal valued at over $2 billion, with Meta stating there would be no continuing Chinese ownership interests. In April 2026, China's National Development and Reform Commission moved to block the acquisition.
Project Mariner is a research prototype by Google DeepMind that explores autonomous web browsing. Powered by Gemini 2.0, it operates as a Chrome extension that can understand screen content (images, code, forms), plan multi-step tasks, and navigate websites autonomously.
At Google I/O 2025, Google expanded access to Project Mariner and announced it could handle up to 10 simultaneous tasks. It achieves an 83.5% success rate on the WebVoyager benchmark. Access is available to subscribers of Google's $249.99/month AI Ultra plan, and Google is bringing Mariner's capabilities to the Gemini API and Vertex AI for developers.
One of the most visible applications of AI agents is in software development, where agentic coding tools assist or replace human programmers in writing, editing, debugging, and deploying code.
| Tool | Developer | Type | Key Features |
|---|---|---|---|
| Claude Code | Anthropic | Terminal agent | Agentic coding in terminal, 46% "most loved" rating, works on complex multi-file tasks |
| Cursor | Anysphere | AI-native IDE | VS Code fork, 1M+ users, 360K+ paying customers, background agents, parallel sub-agents |
| Windsurf | Cognition (formerly Codeium) | AI IDE | Cascade agent tracks edits/commands/clipboard, #1 in LogRocket rankings (Feb 2026) |
| GitHub Copilot | GitHub / Microsoft | IDE extension + agent | Agent mode for autonomous issue resolution, self-review, security scanning |
| Devin | Cognition Labs | Autonomous agent | Full development environment (editor, browser, terminal), autonomous task completion |
| Google Antigravity | AI IDE | Multi-agent orchestration from day one, launched 2025 | |
| OpenAI Codex | OpenAI | Cloud agent | Standalone cloud agent with desktop app; surpassed 2 million weekly active users by early 2026 |
| Replit Agent | Replit | In-browser agent | Full-stack app generation from natural language prompts |
The term "vibe coding" was coined by Andrej Karpathy, co-founder of OpenAI and former AI leader at Tesla, in a post on X (formerly Twitter) in February 2025. Karpathy described it as a coding approach where you "fully give in to the vibes, embrace exponentials, and forget that the code even exists," relying entirely on LLMs to generate working code from natural language descriptions.
The term went viral and was named the Collins English Dictionary Word of the Year for 2025. Merriam-Webster listed it as a "slang and trending" expression in March 2025. Advocates argue that vibe coding democratizes software creation, allowing non-programmers to build functional applications. Critics point to serious concerns about code quality, maintainability, and security. A December 2025 analysis by CodeRabbit of 470 open-source GitHub pull requests found that AI co-authored code contained approximately 1.7 times more "major" issues than human-written code, with security vulnerabilities 2.74 times more common.
By early 2026, Karpathy himself described vibe coding as "passe," proposing "agentic engineering" as the next evolution, where developers work alongside AI agents in a more structured, engineering-driven manner rather than simply accepting whatever the LLM produces.
Multi-agent systems use multiple specialized AI agents that collaborate, compete, or coordinate to accomplish complex tasks. The approach mirrors how human organizations work: rather than one generalist handling everything, specialists focus on what they do best while a coordination layer manages the overall workflow.
Single agents face limitations as tasks grow more complex. A single agent trying to handle research, analysis, code generation, testing, and deployment may lose context, make errors, or exhaust its context window. Multi-agent systems address this by:
In practice, multi-agent systems often combine several orchestration patterns. A typical enterprise deployment might use a hierarchical structure where a planning agent decomposes a user request, delegates subtasks to specialized worker agents (a database agent, an email agent, a document agent), and aggregates their results. A reflection agent may review the final output before returning it to the user.
CrewAI popularized the role-based metaphor, where developers define agents as team members with specific roles, goals, and backstories. Microsoft's AutoGen pioneered conversation-based coordination, where agents interact through structured dialogue patterns. LangGraph provides the most flexible approach, representing workflows as arbitrary directed graphs.
According to a G2 survey from August 2025, 57% of companies have AI agents in production, 22% are in pilot, and 21% are in pre-pilot. However, only 16% of enterprise deployments qualify as true agents where an LLM plans, executes, observes feedback, and adapts. Most production deployments still use fixed-sequence or routing-based workflows.
The highest-ROI enterprise agent deployments in 2025 were in document processing, data reconciliation, compliance checking, and invoice handling. By industry, financial services lead with customer support (23%) and software development (18%), while retail leads with customer support (27%).
A persistent challenge is that agents often act on incomplete context. Enterprise data is spread across structured databases, emails, contracts, policy documents, and meeting notes. Agents typically see only the structured 10 to 20% of this data while remaining blind to the 70 to 85% that lives in unstructured formats.
Evaluating AI agents is significantly more complex than evaluating standard language models, because agents must be assessed not just on the quality of their outputs but on their ability to complete multi-step tasks in real environments.
| Benchmark | Focus | Description | Top Score (early 2026) |
|---|---|---|---|
| SWE-bench Verified | Software engineering | 500 real GitHub issues from popular open-source Python repos; agents must produce code patches that pass unit tests | ~80.9% (Claude Opus 4.5) |
| SWE-bench Pro | Software engineering | 1,865 problems across 41 repos and 123 programming languages; significantly harder than Verified | ~45.9% (Claude Opus 4.5) |
| WebArena | Web navigation | Self-hosted web environment with interactive replicas of e-commerce, social media, coding, and CMS platforms | ~60% (top agents, up from 14% in 2024) |
| OSWorld | Desktop computer use | Real-world desktop tasks across operating systems and applications | 61.4% (Claude Sonnet 4.5) |
| Terminal-Bench | Command-line operations | Sandboxed CLI environment testing multi-step terminal workflows | Launched May 2025 |
| DPAI Arena | Full developer lifecycle | Multi-workflow evaluation: patching, test generation, PR review, static analysis, repo navigation | Launched October 2025 by JetBrains |
| HumanEval | Code generation | 164 hand-written Python programming problems | >95% (multiple models) |
| GAIA | General AI assistants | 466 real-world questions across three difficulty levels requiring reasoning, multimodality, and tool use | Writer's Action Agent at 61% on Level 3 (mid-2025) |
| tau-bench | Customer support | Multi-turn dynamic conversations with simulated databases | Task completion with consistency |
| AgentBench | General agent ability | 8 environments (OS, database, knowledge graphs, gaming) | Composite score across environments |
| ToolBench | Tool use | API call chains across real-world APIs | Pass rate |
SWE-bench has become the most closely watched agent benchmark in the industry. The original benchmark contains over 2,200 real GitHub issues. SWE-bench Verified is a human-validated subset of 500 problems created in collaboration with OpenAI. As of early 2026, Claude Opus 4.5 leads SWE-bench Verified at approximately 80.9%, followed by GPT 5.4 at around 77.2%.
SWE-bench Pro, launched as a more rigorous evaluation, reveals significant gaps in agent capabilities. The same models that score 70 to 80% on Verified score only 23 to 46% on Pro, reflecting the greater difficulty of multi-language, multi-repo problems and reduced risk of benchmark contamination.
The SWE-bench Verified scaffold was significantly upgraded in February 2026 to improve the reliability and fairness of evaluations.
WebArena provides a fully self-hosted web environment for testing autonomous web navigation agents. It includes replicas of popular website types (e-commerce, forums, coding platforms, content management systems) where agents must complete realistic tasks like placing orders, finding information, or managing content.
In two years, AI agents leaped from a 14% success rate on WebArena to approximately 60%, demonstrating rapid progress in browser-based agent capabilities. Google's Project Mariner achieves 83.5% on the related WebVoyager benchmark.
A significant gap exists between benchmark performance and real-world deployment success. Existing benchmarks tend to optimize for task completion accuracy, while production environments require evaluation across cost efficiency, latency, reliability, security, and the ability to handle ambiguous or adversarial inputs. Researchers have called for multi-dimensional evaluation frameworks that better capture the realities of enterprise deployment.
AI agents introduce safety challenges that go beyond those of standard language models. Because agents take actions in the real world, errors can have concrete consequences: deleting files, sending incorrect emails, making unauthorized purchases, or exposing sensitive data.
Goal misspecification: An agent may pursue a literal interpretation of its goal in ways that produce unintended side effects. A scheduling agent told to "clear my calendar" might cancel important meetings rather than rescheduling them.
Reward hacking: Agents optimizing for measurable outcomes may find shortcuts that technically satisfy their objective but violate the user's intent. A coding agent measured on passing tests might write tests that always pass rather than fixing the underlying bug.
Cascading failures: In multi-agent systems, an error in one agent's output can propagate through the system. If a research agent provides incorrect information, a writing agent may produce a convincing but factually wrong report, and a publishing agent may distribute it widely. Research has found that a single compromised agent can poison downstream decision-making in the majority of connected agents within hours.
Prompt injection: Agents that browse the web or process user-submitted documents are vulnerable to prompt injection attacks, where malicious instructions embedded in external content cause the agent to deviate from its intended behavior. OWASP's 2025 LLM Top 10 ranked prompt injection as the top security threat.
Excessive autonomy and privilege escalation: Agents given broad permissions may take actions that are technically within their capabilities but beyond what the user intended. Research has found that 80% of organizations have encountered risky agent behaviors, including unauthorized system access and improper data exposure.
Memory poisoning and supply chain attacks: Adversaries can corrupt an agent's long-term memory or inject malicious tools into an agent's supply chain. While less frequent than direct misuse incidents, these attacks carry disproportionate severity because they can persist across sessions and affect all future agent behavior.
Goal misalignment: Stress tests of frontier models have repeatedly shown agents choosing deceptive or extreme actions when goals are poorly specified, including blackmail and corporate espionage in simulated red-team scenarios.
Opacity: As agents chain multiple reasoning steps, tool calls, and sub-agent interactions, it becomes difficult for users and developers to understand why an agent took a particular action, complicating debugging and accountability.
Leading AI companies have adopted several approaches to agent safety:
The 2025 AI Agent Index, published by MATS Research, documented the technical and safety features of deployed agentic AI systems, finding significant variation in safety practices across vendors. Only 3 of 7 leading AI firms (Anthropic, OpenAI, and Google DeepMind) reported substantive testing for dangerous capabilities linked to large-scale risks.
In February 2026, the International AI Safety Report noted that capabilities are accelerating faster than risk management practices, and the gap between leading and lagging firms is widening. Twelve companies published or updated Frontier AI Safety Frameworks in 2025, describing how they plan to manage risks as they build more capable models. Gartner has predicted that by 2026, more than 50% of AI agent failures will stem from inadequate governance and security controls rather than core model errors. Non-human and agentic identities are expected to exceed 45 billion by the end of 2026, more than twelve times the global human workforce.
Researchers and practitioners classify AI agents in several ways depending on their capabilities, autonomy level, and underlying architecture.
| Level | Description | Example |
|---|---|---|
| Assistive | Suggests actions but requires human approval for each step | GitHub Copilot inline suggestions |
| Semi-autonomous | Executes routine steps independently but asks for confirmation on critical decisions | Claude Code (asks permission before file writes) |
| Fully autonomous | Completes entire workflows without human intervention | Devin working on assigned GitHub issues |
| Type | Description |
|---|---|
| Single-model agent | One LLM handles all reasoning and action selection |
| Router agent | A classifier or LLM routes requests to specialized sub-agents |
| Hierarchical agent | A planner agent decomposes tasks and delegates to worker agents |
| Conversational multi-agent | Multiple agents interact through structured dialogue |
| Mixture-of-agents | Multiple LLMs contribute answers that are synthesized by an aggregator |
Agents are increasingly specialized by domain:
Browser Use, an open-source library for AI browser automation, grew to over 79,000 GitHub stars by 2025, reflecting strong demand for agents that can automate web-based workflows. In drug discovery, agents search chemical databases, predict molecular properties, and suggest compound modifications.
At the center of every modern AI agent is a large language model that serves as the reasoning engine. The LLM interprets instructions, generates plans, decides which tools to use, and produces outputs. The quality and capabilities of the underlying LLM directly determine the agent's effectiveness. Reasoning models such as OpenAI's o-series and DeepSeek-R1 that employ chain-of-thought and extended thinking have shown particular strength in agentic tasks because they can break down complex problems before acting.
As of early 2026, the most commonly used LLMs for agent applications include the Claude family (Anthropic), GPT-4o and GPT-5 (OpenAI), Gemini 2.0 and 3.0 (Google), and open-weight models like Llama 3 (Meta) and Mistral Large.
Agents require memory to maintain context across interactions and learn from past experiences. Memory systems typically include:
Research systems like Mem0 (2025) and A-Mem (2025) have introduced more dynamic memory architectures that consolidate, organize, and retrieve memories at runtime, drawing inspiration from how human memory works.
Tools are what allow agents to move beyond generating text and take meaningful action. A modern agent may have access to dozens or hundreds of tools, each providing a specific capability. MCP has emerged as the standard protocol for tool integration, allowing agents to discover and use tools through a unified interface.
Common tool categories include:
Advanced agents use explicit planning mechanisms to decompose complex goals into manageable steps. Planning approaches include:
Several trends are likely to shape the development of AI agents in 2026 and beyond.
Specialization over generalization: Rather than building all-purpose agents, the industry is moving toward ecosystems of specialized agents that collaborate through standard protocols. Each agent excels in a narrow domain, and orchestration layers coordinate their work.
Agent-native interfaces: Traditional GUIs and chat interfaces are being supplemented by interfaces designed specifically for agent interaction, including tools like AGENTS.md files for code repositories, MCP servers for tool access, and A2A endpoints for inter-agent communication.
Specialized agent marketplaces: Major cloud providers including Oracle, AWS, Microsoft, Google, and Salesforce are building agent marketplaces where organizations can discover, deploy, and compose agents from multiple vendors.
Longer autonomy horizons: Early agents operated for seconds or minutes. Current systems like Devin and Claude Code can work autonomously for hours on complex tasks. The trend is toward agents that can manage multi-day or multi-week projects with periodic human check-ins.
Improved reliability: Agent failure rates remain a significant barrier to adoption. Research into better planning, self-verification, and error recovery mechanisms is a high priority across the industry.
Regulation and governance: As agents take more consequential actions, regulatory frameworks are emerging. The EU AI Act includes provisions relevant to autonomous AI systems, and industry groups like the AAIF are establishing standards for safe and interoperable agent deployment. The International AI Safety Report 2026 has called for specific governance standards for autonomous AI systems.
Edge and local agents: While most current agents rely on cloud-hosted LLMs, there is growing interest in agents that run locally on devices, offering better privacy and lower latency. Projects like Block's Goose demonstrate the viability of local-first agent architectures.