Glossary — Agentic AI

What is an Agent Framework?

1 min read Updated

An agent framework is a software library that provides abstractions for building AI agents, handling concerns like tool management, state persistence, conversation flow, multi-agent coordination, and LLM integration.

WHY IT MATTERS

Building an AI agent from scratch means solving tool calling, state management, error handling, memory, and orchestration yourself. Agent frameworks handle these concerns so developers can focus on business logic.

The landscape has matured rapidly. LangGraph offers stateful graph-based agents. CrewAI specializes in multi-agent teams. OpenAI's Agents SDK provides a minimal, production-focused approach. PydanticAI emphasizes type safety. Each makes different tradeoffs between flexibility and simplicity.

For financial agents, framework choice matters. You need reliable tool execution (failed payment calls must be handled gracefully), state management (agents must track what they've already spent), and extensibility (integrating spending controls into the agent loop).

HOW POLICYLAYER USES THIS

PolicyLayer integrates with popular agent frameworks through SDK plugins and MCP tools. Regardless of which framework you use, PolicyLayer adds a spending control layer that validates financial operations before they execute.

FREQUENTLY ASKED QUESTIONS

Which agent framework should I use?
For complex stateful agents: LangGraph. For multi-agent teams: CrewAI. For production simplicity: OpenAI Agents SDK. For type safety: PydanticAI. For research/experimentation: AutoGen. Start simple and add complexity as needed.
Do I need a framework to build an agent?
No. Many production agents are built with just an LLM API and a while loop. Frameworks help when you need state management, multi-agent coordination, or complex tool orchestration.
How do frameworks handle financial operations?
Most frameworks treat financial tools like any other tool — they don't have built-in spending controls. You need to add a policy layer (like PolicyLayer) to enforce limits on financial tool calls.

FURTHER READING

BUILD WITH POLICYLAYER

Non-custodial spending controls for AI agents. Setup in 5 minutes.

Get Started