Guide
What is agentic AI?
Agentic AI is software that pursues a goal on its own. It decides the next step, uses tools such as web pages, GitHub or Slack, checks what happened, and keeps going until the job is done. An AI agent is one such worker; agentic AI is the approach.
By Abdelhamid NOIRA · Updated
Chatbot or agent: what actually changes
A chatbot waits for a question and gives an answer. An AI agent starts from a goal and works towards it. It can read a page, look at an issue, draft a message and, if you allow it, send that message. The language model is the same kind of model; what changes is that it is given tools and room to take several steps.
| Chatbot | AI agent | |
|---|---|---|
| Starts from | A question | A goal |
| Produces | An answer | A result in the world: a drafted issue, a posted update, a saved note |
| Steps | One reply | Several: plan, act, check, adjust |
| Tools | None, or a few | Web, GitHub, Slack, Notion and more |
| Who stays in charge | You, on every message | You, through rules and approvals |
The five parts of an AI agent
Every useful agent, whatever platform it runs on, is made of the same five things:
- A model. The language model that reads, reasons and writes, such as Anthropic's Claude, OpenAI's GPT models or Google Gemini. Smarter models cost more and take longer, so pick the lightest one that does the job well.
- Instructions. The agent's brief: its role, what it is trying to achieve and how it should work.
- Tools. What it can reach beyond conversation: reading web pages, searching issues, posting to a channel, writing to a notes page.
- Memory. Whether it remembers earlier runs, or starts fresh every time.
- Guardrails. Plain-language rules it must never break, and how much it may do without asking.
In the ANØM console each of these is its own section, and a single sentence is enough to fill all five in as a starting point.
Autonomy: how much should an agent do on its own?
This is the most important setting on any agent, because it decides who is responsible for each action. There are three sensible levels:
- Ask first. The agent pauses for your approval before any tool that changes something, such as posting, writing or creating.
- Balanced. It acts freely on safe steps like reading and searching, and asks before anything irreversible or visible to other people.
- Autonomous. It runs from start to finish and reports back when it is done.
Where agentic AI works well, and where to be careful
Agents earn their keep on repetitive work with a clear outcome:
- Research and monitoring: read a set of public pages and write a short brief.
- Triage: go through new GitHub issues, summarise them and draft replies.
- Routine updates: compose a status message and, after you approve it, post it to Slack.
- Knowledge upkeep: read from and append to Notion pages so notes stay current.
Be more careful with anything involving money, legal commitments, public statements or deleting data. Three habits make agents much safer:
- Give each tool the narrowest access that works, for example a GitHub token limited to a single repository.
- Assume that any text an agent reads from the web, an issue or a message could contain instructions meant to hijack it. This is called prompt injection, and approvals are your main defence.
- Put hard limits on a run: a maximum number of steps, tool calls and minutes.
Build your first AI agent in three steps
- Describe it. Write one sentence, for example “Review new pull requests and flag risky changes.” ANØM chooses a role, a model, tools, rules and an autonomy level, and shows you why.
- Tune it. Change anything in plain English: the model, the tools, the creativity, the rules.
- Dry-run it, then export it. Preview the plan the agent would follow, including where it would stop and ask you, then export a clean JSON or YAML definition. With the ANØM backend connected to your own API keys, the same agent can run for real.
Common questions
How is an AI agent different from workflow automation?
A workflow follows steps you drew in advance and does the same thing every time. An agent decides its own steps, so it copes with messy input, but it is less predictable. Many good setups combine the two: a fixed workflow for the predictable parts and an agent for the judgement calls.
Do I need to write code?
Not to design an agent in ANØM. You describe it in plain English and adjust settings. To run agents against live models you bring your own API keys.
Is agentic AI safe?
It is as safe as the permissions and approvals you give it. Narrow access, human approval for actions that change things, and hard limits per run turn an unpredictable assistant into a dependable one.
Where can I learn about AI agents in Morocco?
Read AI agents in Morocco for how agents work in Arabic, French and English, and what to check before connecting them to customer data. ANØM is made in Morocco by Abdelhamid NOIRA.