A concept reference
for AI agents.
What they are, how they work, the architectural patterns they use, the tooling that surrounds them, and the trade-offs that vendors prefer not to discuss.
A software system that uses a large language model (or similar foundation model) as its reasoning engine to pursue a user-defined goal across multiple steps. Unlike a chatbot, which responds to a single prompt within a conversation, an agent decomposes a goal into sub-tasks, selects and calls external tools (APIs, search, code execution, file systems), observes the results, and revises its plan until it produces an outcome or reports failure.
The term agentic, in current (2025-2026) usage, describes systems in which the language model itself schedules the workflow rather than running as a single step inside a pre-written workflow [1].
Etymology and scope
The phrase software agent has a lineage in 1990s computer science, where it described autonomous processes that perceived an environment, maintained internal state, and acted on goals. Russell and Norvig, in Artificial Intelligence: A Modern Approach (4th ed., Ch. 2), formalised the framework still used in classroom teaching: simple reflex agents, model-based reflex agents, goal-based agents, and utility-based agents [2].
The 2024-2026 surge of interest in agents is narrower. It refers specifically to LLM-based agents: systems where a large language model performs the reasoning that classical agents implemented as rule trees or planning algorithms. The classical taxonomy still applies, but the reasoning substrate is now a foundation model rather than hand-authored logic.
Boundary cases recur. Is GitHub Copilot an agent? Conventionally no, since its review-and-accept design keeps the human in control of every output; it is closer to an assistant. Is an autonomous vehicle an AI agent? Yes in the classical sense, no in the LLM-based sense the term carries today. Is a fixed pipeline that calls an LLM to summarise a document an agent? No; the LLM is a node, not the scheduler.
Four sections, sixteen reference pages.
Numbered as the front matter of a reference volume. Each page answers a single conceptual question, grounds its claims in publicly citable sources, and refuses to recommend products by price.
- I.Foundations
How the perceive-reason-act-evaluate loop works and how agents differ from adjacent categories: assistants, copilots, RPA bots, chatbots, and pre-defined workflow automation.
- II.Patterns
The architectural patterns practitioners actually use, drawn from the research literature and framework documentation. ReAct, planner-executor, reflection, multi-agent orchestration, and the underlying tool-use mechanism.
- III.Tooling
Four categories of infrastructure that surround a working agent: orchestration, observability, evaluation, and memory. Vocabulary references rather than product recommendations.
- IV.Decisions
Decision frameworks written as principles, honest limitations of the technology, an architecture-level build guide, and a pattern-level use-case reference.
An editorial project, not a buyer’s guide.
This site does not rank products, does not publish pricing tables, and does not receive vendor sponsorship. It is a concept reference, maintained as an independent editorial project. Where specific frameworks or platforms are referenced, they are linked to their public documentation for reader convenience, not as recommendations.
The voice is deliberately uncommercial. Where the content cites a number, that number is sourced. Where a pattern is described, the original paper is cited. Where a claim is editorial synthesis rather than established fact, it is marked as such.
Why this matters now.
Agents have been a research topic for decades. The recent surge in production interest follows three concurrent technical maturations, each with a public, verifiable track record:
First, model providers improved the reliability of structured tool calling. Anthropic, OpenAI, and Google all shipped tool-use APIs that emit syntactically valid JSON arguments at rates substantially higher than earlier generations, as documented in their respective model cards and changelogs [3].
Second, inference cost per token continued to fall. The Stanford AI Index Report tracks this trend each year as part of its Technical Performance chapter, and the curve is one of the few in the field that is publicly verifiable without paywalled analyst reports [4].
Third, production-grade orchestration frameworks reached working maturity. LangGraph, Microsoft AutoGen, CrewAI, and the OpenAI Agents SDK each shipped stable releases through 2024-2025, and their public release histories give a verifiable timeline of capability without the need to quote analyst surge statistics [5].
No market-size figure or analyst-surge percentage appears on this site. Where the literature contains a stat that cannot be sourced to a publicly accessible URL, we omit the stat.
Five places to start.
- Anthropic, Building effective agents (2024). The clearest public treatment of the workflow-vs-agent distinction.
- S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson (2020). Ch. 2 remains the canonical reference for the classical agent taxonomy.
- L. Wang et al., A Survey on Large Language Model based Autonomous Agents, arXiv:2308.11432 (2023). The most-cited LLM-agent survey.
- Stanford HAI, AI Index Report (annual). Technical performance, cost, and adoption data sourced from publicly available datasets.
- S. Yao et al., ReAct: Synergizing Reasoning and Acting in Language Models, arXiv:2210.03629 (2022). The foundational ReAct paper.
Sources and Further Reading
- Anthropic, Building effective agents (December 2024).
- S. Russell and P. Norvig, Artificial Intelligence: A Modern Approach, 4th ed., Pearson, 2020.
- OpenAI function-calling guide; Anthropic tool-use documentation; Google Gemini function-calling guide.
- Stanford HAI, AI Index Report (annual).
- LangGraph documentation, AutoGen documentation, CrewAI documentation, OpenAI Agents SDK repository.
- L. Wang et al., A Survey on LLM-based Autonomous Agents, arXiv:2308.11432 (2023).
- S. Yao et al., ReAct: Synergizing Reasoning and Acting, arXiv:2210.03629 (2022).