Tools & agents
AI Agent
Also known as: agentic AI
An AI system that can take actions to reach a goal, using tools, taking steps and reacting to results, not just producing a single answer.
An AI agent is an AI that does not just answer, it acts. Instead of only telling you how to do something, it can take the steps itself: search the web, fill in a form, send an email, check the result and try again if needed. Think of the difference between an assistant who gives you directions and one who actually drives you there.
An AI agent uses a language model as its "brain" but adds the ability to act in a loop: decide what to do, use a tool, look at the result, and decide the next step, repeating until the goal is reached. A plain chatbot answers once; an agent works through a task.
What makes something an agent
- Tools: it can call external functions, search, code execution, sending messages, databases.
- A loop: it plans, acts, observes the outcome, and adjusts.
- A goal: it works toward an objective rather than a single reply.
Standards like the Model Context Protocol give agents a uniform way to connect to those tools, and techniques like RAG let them pull in the right information.
An AI agent couples an LLM policy with an action space and an environment, running an iterative perceive-reason-act loop rather than a single forward generation. A common pattern interleaves reasoning traces with tool calls, ReAct-style, so the model plans, invokes a tool, observes the result, and conditions subsequent steps on it (Yao et al. 2022). Building blocks include tool/function calling, memory (short-term via the context window, longer-term via retrieval), planning/decomposition, and often multi-agent orchestration.
In practice, reliability, not capability, is the hard part: error compounding across steps, cost and latency of long loops, and safe handling of side-effecting actions. Guidance from practitioners emphasises preferring the simplest workable design, composable and well-scoped tools, explicit stopping conditions, and human checkpoints for irreversible actions (Anthropic, 2024). Interoperability layers such as MCP standardise how agents discover and call external capabilities.