Agentic Levels

Wissen

Not every "agent" is equally autonomous. The Agentic Levels describe how much decision-making freedom a system has. This distinction matters because it directly affects which architecture and which guardrails you need.

Hover or clickTap a level to see details

Level 1: Output Decisions

The agent decides on the format and content of its response, but not the process. It receives an input and generates an output -- no tools, no loops, no planning.

Examples: Chatbot (brief vs. detailed), email classifier, translator with tone adaptation.

Input → [LLM] → Output

Guardrails: Output validation, toxicity filters, fact-checking for critical information.

Level 2: Task Decisions

The agent decides which steps to take and which tools to use. It has a goal and independently finds its way -- this is the classic ReAct agent.

Examples: Research agent, coding agent, support agent with database access.

Input → [LLM + Tools] → Loop → Output
              ↑            │
              └────────────┘

Guardrails: Everything from Level 1, plus max-steps limit, tool permissions (least privilege), human-in-the-loop, logging.

Level 3: Process Decisions

The agent doesn't just decide on steps, but on the working method itself: Which other agents should be deployed? How many instances? In what configuration? This is the level of multi-agent systems.

Examples: Orchestrator for coding projects, research coordinator with dynamic agent count, incident response system.

Input → [Orchestrator-LLM]
              │
              ├── Spawn: Agent A (Model X, Tools Y)
              ├── Spawn: Agent B x3 (parallel)
              └── Spawn: Agent C (waits for A+B)
                         │
                         └── Results → [Orchestrator] → Output

Guardrails: Everything from Level 1 and 2, plus budget limits, agent spawn limits, timeout, monitoring/alerting, kill switch.

!With Great Autonomy Comes Great Responsibility

Level 3 agents can incur significant costs in a short time and take unexpected actions. Each additional level of autonomy requires exponentially more guardrails. Always start with the lowest level that your task requires.

Verstehen

How to Choose the Right Level

Question	Level 1	Level 2	Level 3
Does the agent need tools?	No	Yes	Yes
Are the steps predictable?	Yes	Partially	No
Do multiple agents need to cooperate?	No	No	Yes
How high is the risk?	Low	Medium	High
How much control do you need?	Little	Medium	A lot

Real-World Example: E-Commerce Platform

Consider an e-commerce platform and how different features operate at different levels:

Level 1: Generating product descriptions -- Input: product data, Output: text. No agentic behavior needed.
Level 2: Support agent -- answers customer questions, looks up orders, initiates returns. Classic ReAct agent with tools.
Level 3: Pricing optimization -- an orchestrator analyzes market data (Agent A), competitor prices (Agent B), inventory levels (Agent C), and calculates optimal prices. The number and type of analysis agents depend on the product category.

Cross-Level Architecture

In real systems, you often work with multiple levels simultaneously. A Level 3 orchestrator delegates to Level 2 agents, which make Level 1 calls to LLMs. The art lies in equipping each level with the right guardrails and restricting autonomy to the minimum that the task requires.

Level 3: Orchestrator (Process Decisions)
    │
    ├── Level 2: Research-Agent (Task Decisions)
    │       └── Level 1: Summarizer (Output Decision)
    │
    ├── Level 2: Analysis-Agent (Task Decisions)
    │       └── Level 1: Formatter (Output Decision)
    │
    └── Level 2: Writer-Agent (Task Decisions)
            └── Level 1: Translator (Output Decision)

Each level adds autonomy -- and requires correspondingly more safeguards.

Anwenden

Your company wants a system that automatically analyzes incident reports and responds differently based on severity: for critical incidents, 5 diagnostic agents should run in parallel; for normal incidents, one is enough. Which Agentic Level describes this system?

Reflect

Agentic levels help you consciously control the degree of autonomy in an agent system. The higher the level, the more decision-making freedom the agent has -- but also the more safeguards you need. In the next section, you will learn the best practices that help you run multi-agent systems reliably and cost-effectively.