Agentic AI
Aussi : AI Agents, Autonomous AI Agents, Agent-based AI
Agentic AI refers to AI systems that pursue goals autonomously by planning, taking actions through tools, and adapting based on results, with minimal step-by-step human direction.
What It Is
Agentic AI describes AI systems that act as autonomous agents: they receive a goal, break it into steps, decide which actions to take, use external tools or APIs, observe the results, and adjust their approach until the goal is met. Unlike a single prompt-and-response model, an agentic system runs in a loop, reasoning about what to do next.
Most current implementations wrap a large language model (LLM) with three additional capabilities:
- Planning: decomposing a high level goal into ordered tasks.
- Tool use: calling functions, APIs, databases, search engines, or code execution.
- Memory: retaining context across steps and sometimes across sessions.
Why it matters
Traditional automation follows fixed, hard-coded rules. Agentic AI handles tasks that are open-ended or change with circumstances, choosing actions dynamically instead of following a rigid script. This matters because it shifts AI from answering questions to completing multi-step work: researching, comparing options, executing transactions, and reporting back.
The trade-off is control. More autonomy means more risk of errors, unexpected actions, or runaway loops, so guardrails, approval steps, and observability become essential.
How it is used in practice
- Data: an agent queries a warehouse, runs validation checks, and drafts a summary of anomalies.
- Marketing: an agent researches competitors, generates campaign variants, and schedules them across channels.
- Finance: an agent reconciles invoices against purchase orders and flags mismatches for review.
- Software: an agent reads a ticket, edits code, runs tests, and opens a pull request.
Common patterns include a single agent with tools, multi-agent setups where specialized agents collaborate, and human-in-the-loop designs that require approval before sensitive actions.
A concrete example
Goal: "Book the cheapest flight to Berlin next Tuesday under 300 euros."
1. The agent plans: search flights, filter by price and date, then book.
2. It calls a flight search API (tool use).
3. It observes results, finds none under 300 euros, and widens the time window (adaptation).
4. It finds a match and pauses for human confirmation before paying.
5. It completes the booking and reports the itinerary.
This loop of reason, act, observe, and repeat is the defining characteristic of agentic AI.