Article 1/4 in the AI Agents series — How to build an AI agent? (coming soon) · AI agents in the team (coming soon) · AI agents vs. organizations (coming soon)
ChatGPT answers your questions. An AI agent carries out missions. Behind this shift in vocabulary lies a technical breakthrough that will redefine what we expect from software — and from the people who steer it.
"What's the difference between an AI agent and a skill?"
The question came up on a Friday afternoon during a team sync. Why does Claude [Ed. note: Claude Code] talk about "skills"? Why does everyone else talk about "agents"? How do they work together?
It's a great question. And the fact that it's hard to answer in thirty seconds says a lot about the state of AI vocabulary in 2026.
The answer fits in a simple metaphor: the agent is the carpenter, the skill is its hammer.
A skill (or "tool") is a specific, bounded capability. It's a mechanism that lets the AI model interact with its environment: reading a file, running a web search, or executing a SQL query.
An agent, on the other hand, is the autonomous system that decides when and how to use these skills. It receives a goal, plans the necessary steps, and calls the right skill at the right time to move forward.
The Claude Code example illustrates this complexity well. Its architecture stacks four levels: a main orchestrator agent, skills (complex action programs), specialized sub-agents that those skills can mobilize, and atomic tools (read a file, run a search). What another vendor would call an "agent", Claude Code calls either a skill or a sub-agent — depending on the level of autonomy granted.
Even though the industry still produces jargon at industrial pace — what one vendor calls a tool, another calls a skill — this "brain that decides / hands that execute" duo remains the core principle. You'll find all this vocabulary detailed in the site glossary. But for now, let's look at what this "brain" is made of under the hood.
Definition: what makes a system an "agent"
An AI agent is a software system capable of perceiving its environment, reasoning about a goal, and executing actions autonomously — in a loop, until it reaches that goal.
Three words matter in this definition: perceive, reason, act. This triptych fundamentally distinguishes an agent from a simple LLM. ChatGPT, Claude, or Gemini in their basic usage are question-answer systems: you enter a prompt, you get an output, the interaction stops. An agent orchestrates multiple steps, decides its own intermediate actions, and corrects course based on the results obtained.
The distinction comes down to one word: autonomy. An LLM answers. An agent decides.
In technical terms, to go from this conceptual triptych to a working agent, you build the system around an LLM (which serves as the "brain") by adding four software building blocks:
1. Memory (to perceive over time) — storage of past interactions (short-term memory) and persistent knowledge (long-term memory, often vector-based). Without memory, an agent forgets everything between executions and can't build on its past actions.
2. Planning (to reason) — the ability to break down a complex goal into sequential or parallel sub-tasks. The agent doesn't receive a script to follow: it builds its own action plan based on context. Without planning, you'd have to script every sequence by hand — and everything breaks as soon as the context changes.
3. Tools (to be able to act) — functions the agent can call: external APIs, web search, code execution, file read/write, SQL queries… This is the operational arm. An agent without tools is a brain without hands.
4. Action (to interact with the real world) — actual execution of the planned sub-tasks through the available tools. This is the layer that turns reasoning into a concrete result: an email sent, a file created, a CRM updated.
AI agent vs. chatbot: where does the real breakthrough lie?
Confusing the two is costly — in miscalibrated expectations, disappointments, and sometimes abandoned projects after underwhelming results.
A chatbot — even a sophisticated one, even LLM-based — is fundamentally reactive and stateless: each exchange is processed independently. It takes no initiative, consults no external system unless explicitly configured, and remembers nothing between sessions.
An AI agent is proactive and stateful: it maintains context between steps, decides its own next actions, and iterates until it succeeds — or fails in a controlled manner. Given the goal "Prepare a competitive brief on our three main suppliers", it autonomously chains: web search → data extraction → structuring → writing → PDF export — with no human intervention between steps.
| Criterion | Chatbot | AI Agent |
|---|---|---|
| Mode | Reactive, turn-by-turn | Proactive, multi-step |
| State | Stateless (or limited memory) | Stateful, persistent |
| External tools | Rare or absent | Native |
| Planning | None | Autonomous |
| Error handling | No automatic retry | Retry, branching, escalation |
| Human intervention | At every exchange | Optional (human in the loop) |
The boundary isn't always clear-cut in commercial products — Claude, Gemini, and GPT-4 can adopt agentic behavior depending on their configuration — but the architectural distinction holds.
How it works: the ReAct loop
An agent's operation relies on an iterative loop known as ReAct (Reasoning + Acting), formalized by Yao et al. in 2022 and now the foundation of most modern frameworks.
1. Perceive — The agent receives a goal and observes the current state of its environment: available data, history of past actions, results from previous steps.
2. Reason — The LLM analyzes the situation and produces an action plan: which sub-task to prioritize, which tool to call, how to formulate the request. This is where the quality of the underlying model makes the difference.
3. Act — The agent executes its chosen action — API call, web search, Python script, file read — and awaits the result. This is the only step that produces an effect in the real world.
4. Observe — The agent analyzes the result, compares it to the goal, and decides: continue (new iteration), correct (change strategy), or terminate (goal reached or unrecoverable error).
This cycle can repeat dozens of times for a single complex mission. This "looping" behavior is exactly what the term agentic AI covers — an AI that acts, not an AI that answers.
Multi-agent architectures: divide and conquer
A single agent efficiently handles sequential tasks. For complex workflows — those requiring varied skills or parallel processing — you orchestrate multiple specialized agents: this is the multi-agent architecture.
The fundamental pattern pits two complementary roles against each other.
The orchestrator is the "project manager" agent: it receives the overall goal, breaks it down into sub-tasks, delegates them to the right specialized agents, and aggregates the results. It's the only one that interacts with the end user. It doesn't necessarily know how to do things — it knows who does them and in what order.
Sub-agents are the expert agents: each masters a specific domain — web search, data analysis, writing, regulatory validation, specific API calls. They receive a sub-task, execute it, and return a structured result to the orchestrator.
More advanced architectures add parallelism (multiple sub-agents simultaneously process independent tasks) and the critic pattern (one agent evaluates another's output before validation).
On a content platform, the architecture could deploy five specialized agents: SEO research (Ahrefs/Rankplorer for French), factual research with domain filtering (Perplexity API or Brave API), writing (Mistral Large), anti-hallucination verification (Mistral Magistral), and final formatting. The orchestrator coordinates the chain based on the type of content to produce. In an internal test of 50 articles, the pipeline produced editorial quality deemed publishable without major revision, and no hallucinations detected during human review on regulatory data and cited sources.
Agentic RAG: when the knowledge base becomes active
Classic RAG works in two stages: the agent queries a document base, retrieves relevant passages, and injects them into the LLM's context to generate a response. It's effective — but linear and rigid.
Agentic RAG adds intelligence to the retrieval phase: the agent decides how to query the base (query reformulation, parallel queries), whether the results are sufficient, and when to supplement with an external source. Retrieval becomes one action among many in the ReAct cycle, not a fixed upstream step.
In practice:
- Classic RAG: question → single vector search → answer
- Agentic RAG: question → planning → search 1 → evaluation → if insufficient: reformulation + search 2 → possibly web search → synthesis → answer
Quality on complex or ambiguous questions improves significantly — at the cost of higher latency and more complex infrastructure.
Three examples to anchor the concept
Automated regulatory monitoring — Every morning, an analyst opens the Official Journal, scans ministerial circulars, cross-references with industry sources, and hopes not to miss anything. Two hours later, they have a summary — if they haven't been interrupted. An agent does the same in a few minutes: it monitors these sources daily, produces a structured summary as soon as a publication matches predefined criteria, and sends it to the right person. The remaining time goes to review and action.
SEO content production at scale — Manually, a writer analyzes keywords, searches for sources, writes, fact-checks, and formats everything. Result: two to three pieces of content per day, at best. An agent chains these same steps — ranking data, factual sources, writing, verification, HTML formatting — and processes dozens of pieces per hour. Important nuance: Google devalues mass-generated content without editorial added value. The agent accelerates production, but without human review and a unique angle, organic traffic eventually plateaus or declines.
B2B lead qualification — A new contact arrives in the CRM. The sales rep opens LinkedIn, company registries, searches for ongoing tenders, estimates IT budget… Twenty minutes per lead — before even picking up the phone. An agent does this work in seconds: it analyzes the incoming contact, searches public information (size, IT budget, recent communications, ongoing tenders), evaluates the fit against qualification criteria, and enriches the CRM with a structured profile ready to use.
In all three cases, the common thread is identical: a defined mission → a ReAct cycle → real tools → a result delivered without human intervention at each step.
What this article doesn't cover — and why
This article answers a specific question: what is an AI agent? It intentionally doesn't address the frameworks to build one, their deployment within teams, or their impact on organizations. Each of these topics deserves its own treatment.
The full series:
- Article 1 (this one) — What is an AI agent? Definition, architecture, examples
- Article 2 — How to build an AI agent? Frameworks, steps, and pitfalls to avoid (coming soon)
- Article 3 — AI agents in the team: concrete use cases and first deployments (coming soon)
- Article 4 — AI agents vs. organizations: governance, autonomy, and strategic questions (coming soon)
Article 2 will get practical: how to build an agent, which frameworks to choose, and the pitfalls to avoid. To get it as soon as it's published, subscribe to the RSS feed or follow me on LinkedIn.
