What are AI Agents?

Abstract

This NVIDIA glossary article provides a comprehensive primer on autonomous AI agents — systems that reason, plan, and execute multi-step tasks based on high-level goals within security, privacy, and policy constraints. The article defines the five core components of an AI agent (LLM reasoning brain, memory modules, planning modules, tool integrations, and systems of models), walks through a step-by-step execution example (sales data analysis), enumerates seven agent types from simple reflex to utility-based, surveys four orchestration patterns (centralized, decentralized, federated, hierarchical), draws a comparative table between agents and AI assistants, and catalogs representative use cases including task execution, workflow optimization, data analysis, customer service, software development, and supply chain management. NVIDIA positions its Blueprints, API catalog, and NVIDIA OpenShell runtime as primary platforms for developing and deploying agents at scale.


Key Concepts

  • Autonomous AI Agent: An LLM-powered system that reasons, plans, and executes multi-step tasks with minimal human input, operating within security, privacy, and policy constraints defined by the deployment environment. Contrasted with generative AI’s “request-and-respond” model — agents orchestrate and collaborate with other agents and tools to handle complex workflows.
  • Five Core Components:
    1. LLM — decision-making brain; coordinates reasoning, planning, tool selection, and data access within guardrail constraints
    2. Memory Modules — short-term (current workflow context) and long-term (historical knowledge, preferences) memory
    3. Planning Modules — without feedback: Chain of Thought (CoT), Tree of Thought (ToT); with feedback: ReAct, Reflexion, human-in-the-loop
    4. Tools — APIs (real-time data, programmatic actions), databases and RAG pipelines (knowledge retrieval)
    5. Systems of Models — open models (NVIDIA Nemotron) and frontier models working together for accuracy, cost control, and data privacy
  • Agent Types:
TypeKey CharacteristicsExample
Simple ReflexPredefined rules, no memoryThermostat
Model-Based ReflexShort-term memory, rule-guidedNavigation with traffic updates
Goal-BasedOptimizes path to goalDelivery robot route planning
HierarchicalHigher-level agents manage specialistsFactory automation
LearningAdapts via feedback and experienceAI recommendation system
Multi-Agent Systems (MAS)Coordinated agents achieving common goalsDrone delivery fleet
Utility-BasedMaximizes utility/reward per actionDynamic pricing algorithm
  • Orchestration Patterns:
PatternDescriptionStrengthsChallenges
CentralizedSingle supervisor agent coordinates all tasksClear control, consistent decisionsBottlenecks, less adaptable
DecentralizedEach agent operates autonomouslyHigh flexibilityComplex communication protocols
FederatedMultiple agent systems collaborate across orgsCross-system collaborationRequires shared standards
HierarchicalTiered supervision structureBalances flexibility and oversightCoordination complexity
  • Agents vs AI Assistants: Assistants (Siri, Alexa) are reactive, low-autonomy, minimal learning, limited integration. Agents are proactive, high-autonomy, learning-capable from interactions, and extensively integrated with APIs, databases, and tools — enabling iterative, open-ended workflows.
  • NVIDIA OpenShell: Open source sandbox runtime for safe long-running agents — zero permissions by default, private inference, every action policy-enforced at infrastructure layer, no code changes required.
  • Guardrails at Infrastructure Layer: Sandboxes, policy engines, and privacy routers manage tool access and protect sensitive data. The article explicitly distinguishes infrastructure-layer enforcement from guardrails inside agent code — the former is required for enterprise production because agents run for long periods and access both online and local data.

Key Claims and Findings

  • Autonomous AI agents represent the “next evolution” of AI — transitioning from simple automation to systems capable of managing complex workflows within security and governance boundaries.
  • The reasoning layer is the defining feature of agentic AI: by combining LLM capabilities with tools, orchestration software, and contextual memory, agents navigate complex environments with precision and adaptability.
  • Enterprise agents require guardrails at the infrastructure layer, not just inside agent code — because agents run for long periods and access both online and local data, sandboxing, policy engines, and privacy routers are non-negotiable for production deployment.
  • AI agents and AI assistants differ structurally, not just in degree: assistants are reactive and command-driven; agents are proactive, iterative, and self-correcting based on feedback and prior interactions.
  • Orchestration is foundational for scalability: it enables dynamic problem-solving (adapting to changing conditions), improved resource utilization, and enhanced system reliability across complex multi-agent deployments.

Terminology

  • Chain of Thought (CoT): Structured technique to decompose tasks into manageable steps without external feedback — the agent reasons through intermediate steps before acting.
  • Tree of Thought (ToT): Branching decomposition that explores multiple solution paths simultaneously, selecting the most promising branch.
  • ReAct: Interleaves reasoning and action — the agent reasons about what to do, takes an action, observes the result, then reasons again for iterative refinement.
  • Reflexion: Post-action self-reflection — the agent evaluates the effectiveness of its past actions to improve future strategy.
  • OODA Loop (Observe-Orient-Decide-Act): Control loop referenced in the data-center optimization use case, used as an agent strategy for continuous environmental adaptation.
  • API Agent / Execution Agent: A task execution agent that carries out user-requested tasks via a set of predefined executive functions (tool calls to external APIs).
  • NVIDIA OpenShell: Open source agent runtime providing zero-trust sandboxing; agents start with zero permissions and earn access through policy enforcement.

Use Cases

DomainAgent Role
Task execution”API agent” calls predefined functions to complete user-specified tasks (e.g. write and distribute social media content)
Workflow optimizationAI co-pilots help users understand application features and automate their use; data-center OODA-loop swarm agents
Data analysisMulti-agent “extract and execute” pattern: one agent set gathers data from memory/PDF, another calls analysis APIs
Customer service24/7 NLP-driven support that connects to CRM, checks refund eligibility, inputs return data
Software developmentGitHub Copilot: code suggestions, error fixes, PR summaries, documentation generation
Supply chain managementHierarchical swarm monitors inventory, demand, and raw material markets; reports to orchestrating supervisor agent

Connections to Existing Wiki Pages