Agent Capability Patterns
Capability patterns are the architectural building blocks that make AI agents powerful. While a basic LLM generates text in a single pass, an agentic system combines multiple patterns — reflection, tool use, planning, collaboration, and memory — to tackle complex, multi-step tasks autonomously.
Four of these patterns, Reflection, Tool Use, Planning and Multi-Agent Collaboration, are the agentic design patterns Andrew Ng set out in a five-part series in The Batch (DeepLearning.AI) in March and April 2024, each with suggested research papers (Part 1, March 20, 2024). The playbook adds three more: Memory, and two patterns for safety and control. The patterns are platform-agnostic — they apply whether you’re building with Claude, ChatGPT, Gemini, or any other LLM.
Pattern Catalog
Section titled “Pattern Catalog”Capability Patterns
Section titled “Capability Patterns”These five patterns define what agents can do. The first four are Ng’s; Memory is the playbook’s addition.
| Pattern | What It Does | Key Benefit |
|---|---|---|
| Reflection | Agent reviews and critiques its own output, then improves it | Higher-quality results through self-correction |
| Tool Use | Agent calls external tools, APIs, and data sources | Extends capabilities beyond text generation |
| Planning | Agent breaks complex goals into a sequence of steps | Handles multi-step tasks that require strategy |
| Multi-Agent Collaboration | Multiple specialized agents work together on a task | Tackles problems too complex for a single agent |
| Memory | Agent stores and retrieves information across interactions | Learns from experience and maintains context |
Safety & Control
Section titled “Safety & Control”These two patterns define how agents stay safe and aligned.
| Pattern | What It Does | Key Benefit |
|---|---|---|
| Guardrails | Automated rules that constrain agent behavior | Prevents harmful or off-topic outputs without human intervention |
| Human Checkpoint (Human-in-the-Loop) | Human checkpoints at key decision points | Keeps humans in control of high-stakes actions |
How Patterns Work Together
Section titled “How Patterns Work Together”Consider a customer exchange request — a customer wants to return a product and exchange it for a different item. A capable agent doesn’t just use one pattern; it combines several:
- Planning — The agent breaks the exchange into steps: verify the order, check return eligibility, find the replacement item, process the return, place the new order.
- Tool Use — At each step, the agent calls the order management API, inventory system, and payment processor.
- Reflection — Before confirming the exchange, the agent reviews its work: “Did I verify the return window? Is the replacement item in stock? Are the prices correct?”
- Guardrails — Automated rules prevent the agent from issuing refunds above a threshold or processing exchanges for final-sale items.
- Human-in-the-Loop — If the exchange involves an exception (expired return window, high-value item), the agent escalates to a human agent for approval.
- Memory — The agent remembers this customer’s preferences and past interactions, personalizing future service.
No single pattern makes this workflow possible. Their combination is what turns a basic chatbot into a capable agent.
Where to Start
Section titled “Where to Start”Understanding agent concepts? Start with Reflection — it’s the simplest pattern to grasp and demonstrates the core idea of iterative improvement.
Building your first agent? Tool Use is the most immediately practical pattern — it’s how agents interact with the real world.
Designing a production system? Read Guardrails and Human Checkpoint first — safety and control should be designed in from the start, not bolted on later.
Related
Section titled “Related”- Agents — The Agents building block overview
- Agents Resources — Recommended reading on agents
- Workflow Architecture Patterns — how the work is structured, mapped to autonomy levels
- Building Effective Agents — Anthropic’s guide to agent architecture
- Prompt Engineering — Techniques for the prompts that drive agent behavior
- Automation Use Cases — Real-world applications of agentic systems