Core concept Confirmed

What is an AI agent?

An AI agent is a system that uses a large language model (LLM) to reason about a task, decide what actions to take, execute those actions (typically via tools), observe the results, and iterate until the task is complete.

Hub: agents Version 1 Reviewed: 2026-07-13 By: editorial

Why this subject matters

AI agents represent a fundamental shift from passive Q&A to active task execution. Understanding what agents are — and what they are not — is critical for anyone building, buying, or evaluating AI-powered systems.

Current status

Agent architectures are maturing rapidly. The basic pattern (LLM + tools + loop) is well-established, but production-grade reliability, safety, and observability remain active areas of research.

Core concepts

Agent loop
The core execution pattern: (1) observe current state, (2) reason about next action, (3) execute action via tool, (4) observe result, (5) repeat until goal is reached. Sometimes called the 'sense-reason-act' loop.
Tool use
The mechanism by which an agent interacts with the world — calling APIs, reading files, executing code, querying databases. Tools are the agent's hands.
Planning
The agent's ability to decompose a complex goal into sub-tasks, order them, and adapt the plan as new information arrives.
Memory
Agents need to remember: conversation history, facts from tools, decisions and outcomes, and user preferences. Memory can be short-term (in-context) or long-term (persisted).
Orchestration
The layer that manages agent execution — deciding which agent handles what, managing retries and error recovery, enforcing timeouts, and coordinating multi-agent interactions.

What it does

AI agents can: execute multi-step tasks autonomously, use tools to interact with external systems, adapt their plan when new information changes the situation, ask clarifying questions, and operate in loops until a task succeeds.

What it does not do

AI agents do not: have consciousness or genuine understanding, always succeed, replace human judgment for high-stakes decisions, guarantee safety by themselves, or work well without clear goals and termination conditions.

Known limitations

  • Agents can get stuck in loops, repeatedly calling the same tool without making progress.
  • Goal drift: the agent may pursue a related but incorrect goal if the prompt is ambiguous.
  • Tool failures can derail an agent if error handling is insufficient.
  • Long-running agents consume significant context window space, degrading reasoning over time.
  • Agent behaviour can be non-deterministic; the same prompt may produce different tool-call sequences.

Security and governance considerations

Agents amplify both the capabilities and the risks of LLMs. Key governance requirements: human approval gates for high-impact actions, tool-use policies defining access per circumstance, execution sandboxing, audit logging of every tool invocation, and budget/rate limiting.

Comparisons

Aspect Option A Option B
Control model Chatbot: user drives the conversation Agent: agent drives the execution
Tool access Chatbot: no tool access (or single-shot) Agent: multi-tool, multi-step, iterative
Failure mode Chatbot: wrong answer Agent: wrong action with real consequences

Claims and Evidence

Each material claim below is linked to its evidence source and relationship type. Claims marked "Inference" are editorial synthesis and should not be read as directly sourced facts.

Supports

The agent loop (sense-reason-act) is the foundational pattern for all current AI agent architectures.

Inference Editorial synthesis moderate
Supports

Human-in-the-loop approval is essential for high-stakes agent actions.

Inference Editorial synthesis moderate

Open questions / Disputed areas

  • At what level of reliability do agents become safe to deploy without human-in-the-loop?
  • How should agent memory be structured — single store, per-agent stores, or shared graph?
  • When do multi-agent systems outperform single-agent systems?

Related knowledge pages

Version history (1 revision)
v1 2026-07-13 Initial publication. Defines AI agents, the agent loop, key components, limitations, and governance requirements.