6.3 What Makes a Good Agent?

Responsive and adaptive AI.

🎯 Core Goals

  • Define what makes an AI system truly “agentic” — not just a script with tools.
  • Understand the six key characteristics of high-quality agents.

A real agent isn’t just an LLM with tools bolted on. It operates autonomously, breaks down complex goals, adapts when things go wrong, and can collaborate — like an employee who figures things out, not just a tool that waits to be used.

👁️ Visuals & Interactives

The Agent Checklist

What separates a smart script from a true AI agent?

self_improvement

Autonomy & Proactivity

Acts independently toward a goal — doesn't wait to be told every step. More like an employee than a tool.

checklist

Goal-Oriented Planning

Breaks high-level goals into actionable steps in a logical order — prerequisites first, then build on them. Prioritizes, sequences, and adjusts when conditions change.

build

Tool Use & Interoperability

Interacts with external APIs, search engines, and databases to get real things done — not just talk about them.

published_with_changes

Adaptability

If Tool A fails, it tries Tool B. Uses real-time feedback to refine its approach mid-task.

memory

Context & Memory

Maintains short- and long-term memory of past steps, enabling coherent work across multi-step or multi-session tasks.

group

Collaboration

Works alongside humans and other specialized agents to accomplish complex, multi-part projects.

Script
A → B → C

Fixed steps, breaks if anything changes

Agent
Perceive → Plan → Act → Observe → Repeat

Adaptive loop — keeps going until the goal is met

📝 Key Characteristics

  • Autonomy & Proactivity: Unlike passive, reactive AI, a good agent acts independently to achieve goals. It doesn’t wait to be told every step — it decides what to do next.
  • Goal-Oriented Planning: Agents break complex, high-level goals into actionable steps, prioritize them, and adjust strategies when conditions change.
  • Tool Use & Interoperability: Good agents interact with external systems — APIs, software, search engines, databases — to get tasks done (e.g., checking email, updating a CRM).
  • Adaptability & Learning: Agents use feedback loops and real-time data to refine their performance and adapt to new situations mid-task.
  • Context Awareness & Memory: Agentic LLMs maintain both short-term and long-term memory of past interactions, allowing for coherent work across multiple steps or sessions.
  • Collaboration: Agents can work alongside humans and often with other specialized agents to accomplish complex, multi-part projects.

A traditional script follows fixed steps: A → B → C. An agent follows an adaptive loop: Perceive → Plan → Act → Observe → Repeat. The loop continues until the goal is met — or the agent asks for help.

Remember those absurd questions, like “Peter bought 12 watermelons, Mary bought 24 apples — how many pineapples does Jack have?” A good model should respond, “No, that’s impossible.” Similarly, if you ask an impossible question, a good LLM shouldn’t say, “What a great idea!” — it should push back. Researchers have created a benchmark to test this capability: Bullshit Benchmark

arrow_back Next arrow_forward