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.
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
The agent loop (sense-reason-act) is the foundational pattern for all current AI agent architectures.
Human-in-the-loop approval is essential for high-stakes agent actions.
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?