Agentic System

System agents published

Also known as: Agentic Application, Agentic AI, Agent System

Definition

A software system that exhibits agent-like behavior—perceiving, reasoning, and acting—but may not fit the strict definition of a single coherent agent. Agentic systems often combine multiple components (LLMs, tools, workflows, retrieval) that collectively produce autonomous-seeming behavior. The term deliberately avoids the philosophical baggage of "agent" while acknowledging the system does more than passively respond.

What this is NOT

  • Not synonymous with 'agent' (agentic systems may lack unified agency)
  • Not just automation (requires reasoning/adaptation, not just rule execution)
  • Not necessarily a single component (often describes composed systems)

Alternative Interpretations

Different communities use this term differently:

llm-practitioners

A system built on LLMs that goes beyond simple query-response patterns to include tool use, multi-step reasoning, or orchestrated workflows that adapt based on intermediate results.

Sources: Anthropic: Building Effective Agents (2024), Industry usage in 2023-2024 technical blogs

enterprise-ai

Any AI-powered system that can autonomously perform business processes or workflows, typically with guardrails and human oversight.

Often used interchangeably with "agent" in enterprise contexts, but implies more structured, governed automation.

Sources: Enterprise AI vendor documentation

Examples

  • A customer service system that routes queries, retrieves context, generates responses, and escalates to humans when uncertain
  • A code review system that analyzes PRs, runs static analysis, generates suggestions, and tracks resolutions
  • A research assistant that searches, summarizes, compares sources, and answers follow-up questions

Counterexamples

Things that might seem like Agentic System but are not:

  • A chatbot with only a system prompt and no tools (not agentic—no action capability)
  • A traditional RPA bot following hard-coded rules (automation, not agentic)
  • A classification API endpoint (no multi-step reasoning or action)

Relations

  • generalizes agent (Agents are one way to implement agentic behavior)
  • overlapsWith workflow (Agentic systems often include workflow components)
  • overlapsWith multi-agent-system (Multi-agent systems are a type of agentic system)

See Also