Key Characteristics of Agentic Systems
- List the six key characteristics of agentic AI systems
- Explain the autonomy spectrum and where most production agents sit
- Describe why oversight and autonomy are not mutually exclusive
Beyond the Buzzword
Saying something is "agentic" is not very helpful without knowing what properties you're actually claiming. This lesson defines those properties precisely — so you can evaluate any system (product, demo, or thing you're building) against a consistent standard.
1. Autonomy
Autonomous systems function without requiring explicit instructions for every decision. Given a high-level goal — "research this company and summarize the key risks" — an agentic system figures out the steps without being told what each step is.
Autonomy exists on a spectrum. Human-in-the-loop systems pause and ask for approval at critical junctures. Fully autonomous systems operate without checkpoints. Most production agents today sit somewhere in between — with configurable approval gates for irreversible actions.
2. Goal-Directed Behavior
Agents pursue goals, not just prompts. This is a subtle but important distinction. A chatbot satisfies the current prompt. An agent keeps a goal in mind and evaluates whether its actions are moving toward that goal — even across many steps and tool calls.
This goal-directedness is what enables agents to self-correct. If an action doesn't move toward the goal, a well-designed agent recognizes that and tries a different approach.
3. Multi-Step Operation
Agents autonomously initiate, coordinate, and adapt multi-step tasks. Each step is not a separate human prompt — the agent runs the full sequence after receiving the initial goal. This is the operational signature of an agent: it keeps going.
4. Tool Use
Agentic systems can invoke external tools — search engines, code interpreters, databases, APIs, file systems — and interpret the results. Tool use is what moves agents from "reasoning about the world" to "acting in the world." Without tools, an agent can only think; with tools, it can do.
5. Memory and Context
Agents maintain context across steps. This ranges from holding recent tool results in short-term memory to storing facts in a long-term knowledge base that persists across sessions. Without memory, each step would be a blank slate — the agent couldn't build on its own previous work.
6. Adaptability
When the environment changes or an action fails, an agentic system adjusts its approach. This distinguishes agents from rigid automations that stop or fail silently when they hit an unexpected state. A well-designed agent treats unexpected results as new information and reasons from them.
The Oversight Spectrum
It's worth noting that autonomy and oversight are not opposites. Serious researchers (and frameworks like OWASP's LLM security guidelines) emphasize that agentic systems should be designed with explicit decision boundaries — defined points where human review is required, especially before irreversible actions.
A fully autonomous agent that can delete files, send emails, or charge credit cards without any checkpoint is not a well-designed agent. It's a liability. The best agentic systems are autonomous within carefully chosen limits, not autonomous in spite of them.
- The six agentic properties are: autonomy, goal-directed behavior, multi-step operation, tool use, memory, and adaptability
- Autonomy exists on a spectrum from fully human-in-the-loop to fully autonomous — most production systems sit in between
- Agents without defined oversight boundaries are a liability, not a feature; good agentic design includes explicit checkpoints
- Goal-directed behavior is what enables self-correction — the agent evaluates whether its actions are moving toward the goal
Progress is saved in this browser only — clear your cache or switch device and it's gone. Create a free account and it follows you everywhere.