Agentic AI
How it works
The agentic system receives a goal, then independently plans steps, selects tools, gathers data, executes actions, and evaluates intermediate results. In simpler variants, a single agent handles this using tool use; in more advanced configurations, multiple agents collaborate on subtasks within a shared workflow.
Problem solved
Traditional generative models handle single prompts well but struggle with extended tasks that require planning, working memory, tool use, and adaptation to changing context. Agentic AI addresses this by combining reasoning, planning, and action execution.
Key mechanisms
Strengths & limitations
Components
Accepts observations from the environment (user messages, tool results, file contents, API responses) and formats them as context for the base model. This may include RAG retrieval to fetch relevant documents.
Official
Decomposes a high-level goal into a sequence of subgoals or actions. The agent may generate an explicit plan or reason step by step using chain-of-thought.
Official
Stores and retrieves information between steps within a session (short-term memory) and optionally across sessions (long-term memory).
Official
The agent is provided with callable external functions: web search, code execution, database queries, file operations, API calls, and browser control. Tool interfaces are defined through schemas such as JSON Schema, OpenAPI, and MCP.
Official
Evaluates whether the current result meets the success criterion. Triggers a retry, replanning, or loop termination. Corresponds to the evaluator-optimizer pattern described by Anthropic.
Official
In multi-agent systems, it directs sub-agents, assigns tasks, and aggregates results. The orchestrator can be an LLM or a statically coded deterministic controller.
Official
Implementation
Model may invoke tools with fabricated parameters or claim to have performed actions it never actually executed — leading to silent failures in multi-step pipelines.
Without a hard step limit or an effective termination criterion, an agent can loop indefinitely, consuming computational resources and hitting API rate limits.
Malicious instructions embedded in tool outputs (web pages, documents, emails) can hijack agent behavior by impersonating system-level instructions.
Accumulated tool outputs and conversation history can exceed the model's context window, causing earlier steps to be silently truncated.
Agents with access to write-enabled tools (file deletion, email sending, database writes) can cause real-world harm when acting on faulty reasoning.
Using agentic autonomy for deterministic, well-defined tasks introduces latency, unpredictability, and failure modes that a simple workflow would avoid.
Evolution
Russell and Norvig formalize rational agents as entities that perceive their environment and take goal-directed actions. BDI (Belief-Desire-Intention) agent architectures are established.
Yao et al. (2022) propose ReAct — interleaving chain-of-thought reasoning traces with action execution in LLMs, demonstrating that language models can serve as a reasoning engine within tool-augmented agentic loops.
OpenAI introduced function calling in GPT-4 in June 2023. AutoGPT, BabyAGI, and LangChain agent abstractions gained widespread adoption. The term "Agentic AI" entered common industry usage.
Andrew Ng's series of blog posts identifies four fundamental design patterns — Reflection, Tool Use, Planning, and Multi-Agent Collaboration — widely cited as a practical taxonomy of agentic systems.
Anthropic published practical guidelines distinguishing workflows (predefined paths) from agents (model-driven execution) and formalized five compositional patterns: prompt chaining, routing, parallelization, orchestrator-workers, and evaluator-optimizer.
Anthropic publishes MCP as an open standard for connecting LLMs to external tool servers, enabling interoperable agentic ecosystems across providers.
LLM-based planners drive robotic actions through perception-planning-action loops, extending agentic paradigms to physical systems and connecting Agentic AI with real-world motor execution.
Hyperparameters (configurable axes)
The set of external tools available to an agent (web search, code execution, file operations, APIs, browser control). It defines the space of possible actions.
A hard limit on the number of reasoning-action iterations before forced termination. Guards against infinite loops.
Whether the agent relies solely on in-context memory or also on external persistent storage (vector database, key-value store).
Whether the system uses a single agent or a network of specialized agents coordinated by an orchestrator.
Whether and at which steps the agent pauses to await human confirmation before taking irreversible actions.
The maximum number of tokens processed by the underlying LLM in a single call. This limits the amount of accumulated history, tool outputs, and instructions that can fit within a single inference step.
Computational complexity
Time complexity: O(N · C_step). Space complexity: O(L_ctx + S_mem).
AgentBench (Liu et al. 2023) - comprehensive 8-environment benchmark (code, games, web nav, DB, OS): frontier LLMs (GPT-5 Sol, Claude Opus 4.8, Gemini 3 Pro) score 45-65% depending on category. GAIA (Mialon et al. 2023) - real-world knowledge work tasks: state-of-the-art (end 2025) ~75% Level 1, ~55% Level 2, ~40% Level 3 - difficulty levels rising with tool count. SWE-Bench Verified (2024) - real-world GitHub issues: SOTA GPT-5.6 Sol Ultra 64.6% single-agent; Claude Fable 5 80%. OSWorld (Xie et al. 2024) - computer use benchmark: SOTA GPT-5.6 Sol 62.6%; Anthropic Claude Computer Use 40-50%. Terminal-Bench (Anthropic 2024) - CLI task completion: SOTA GPT-5.6 Sol Ultra 91.9%. BrowseComp (OpenAI 2025) - web browsing: SOTA GPT-5.6 Sol Ultra 92.2%. Agents' Last Exam (2026) - long-horizon professional workflows: GPT-5.6 Sol 53.6, Claude Fable 5 40.5. Criticism: benchmarks measure success rate but not cost/time/stability - a real production agent must be cheap and deterministically effective.
Compute bottleneck
Each step of the agent loop requires at least one LLM inference call. Multi-step tasks with long context windows and multiple tool calls multiply latency and computational cost linearly.
Execution paradigm
The execution path is not predetermined — it is determined at runtime through the model's reasoning over accumulated context. Workflows with predefined paths represent a degenerate case.
The base LLM decides at each step which tool to call, whether to continue the loop, delegate a task to a subagent, or terminate — based on the current context and observed results.
Parallelism
Parallelism is achievable when subtasks are independent (e.g., parallel web searches, concurrent subagent execution). Sequential loops are required when each step depends on the results of previous tool calls.
Hardware requirements
Agentic AI is an architectural paradigm, not a specific computational kernel. Hardware requirements are entirely determined by the underlying LLM and tools, not by the agent loop itself.
Sources
AI Agent Architecture — ReAct, Memory, Planning and Multi-Agent Systems
A course for developers and AI engineers who want to understand the architectural foundations of AI agents before working with LangChain in practice. After completing this course, you will confidently apply ReAct, planning and memory patterns and understand how to compose agents into multi-agent systems.
Build AI Agents with LangChain
A practical course on building AI agents with LangChain for learners who already know the basics of working with language models and want to create multi-step agentic applications.
AI Agent Security — Attacks, Jailbreaking, and Defense
A course for developers and AI engineers building agentic systems. You will learn real attack vectors against LLMs and AI agents — prompt injection, jailbreaking, model inversion — and how to design layered defenses: guardrails, sandboxing, and AI firewalls.