Your inbox, upgraded.
Receive weekly engineering insights to level up your leadership approach.
Estimated reading time: 5 minutes
Everyone’s been hearing the whispers about agentic AI, but which frameworks are actually worth the investment?
AI Agents have made incredible strides in recent years and are revolutionizing workflows and processes across different industries. Where before, single-agent architectures were able to work for narrow and well-defined tasks, a recent uptick in multi-agent frameworks has overcome the older generation’s limitations.
These new age systems can work together to complete tasks: each agent focusing on what it’s “good” at. The idea is similar to a team in a workplace: rather than one person attempting every job, you have specialists (e.g., a network engineer agent, an infrastructure reliability agent, a software agent, etc.) collaborating to solve different parts of a problem. By distributing cognition, planning, and execution among several agents, a multi-agent system can tackle problems that no single agent could solve alone.
There are many multi-agent frameworks in the market; here are a few you need to know about.
CrewAI
CrewAI is an open-source framework written in Python designed for orchestrating multiple AI agents. The system is framed as a collaborative “crew,” where each agent has a specific role, working together to complete tasks.
CrewAI has a structured approach for its implementation because it enforces a clear and organized design pattern with definite roles and responsibilities, which makes CrewAI comparatively easier to learn. It can be easily integrated with any of the available LLMs, leveraging LangChain tools. CrewAI also allows developers to create their own custom tools.
CrewAI is suitable for production systems as it is optimized for speed due to parallel task execution (running multiple tasks at the same time) and minimized context switching. It is currently being used by major companies like PwC, AWS, and IBM for their workflows.
More like this
LangGraph
LangGraph is an open-source multi-agent framework that uses graph-based architecture (see Fig.1) to help build complex AI agent workflows.
Fig.1. Example of graph-based architecture. Linkurious Enterprise interface with interdependencies between “Location0” entity and the system.
LangGraph lets you design your AI workflow as a flowchart where nodes on the graph represent agents or functions, and the connections define the flow of information. This approach makes relationships between components explicit, leading to easier debugging and testing as you can visually trace data flows between components.
LangGraph takes time to set up as it requires a comprehensive understanding of graph-based architecture, but it gives you strong reliability, visibility, and control, especially for complex, long-running workflows. Its support for long-term memory stores to retain context across multiple conversational sessions enhances its continuity in complex tasks. Additionally, the human-in-the-loop feature gives granular control to users by allowing them to pause workflows for review, validation, or the addition of more context.
While LangGraph integrates tightly with LangChain and supports major LLMs, this coupling can limit flexibility and risk vendor lock-in.
Autogen
Autogen is one of the first open-source multi-agent frameworks that was developed and released by the Microsoft Research team in September 2023. It is built around a simple idea: multiple agents working together in a conversation. These agents can ask each other questions, provide answers, and collectively work through a problem.
Autogen offers high-level, developer-friendly APIs that work with any LLM. You can easily customize the framework to add different agent personalities and then define the interaction and response formats of agents. You can also allow human input to approve agent decisions or create an interactive session between humans and agents. Finally, AutoGen’s asynchronous architecture allows agents to work simultaneously without waiting for each other to complete tasks, provided they don’t depend on each other’s outputs.
OpenAI Agents SDK
In 2024, OpenAI released an experimental multi-agent framework called Swarm for exploration and educational purposes. However, given the product’s infancy, it lacked reliability, granular control, and guardrails for safety, failing to make it to production level.
Eventually, OpenAI moved to a production-ready framework called Agents SDK. It has three definite features: agents, handoffs, and guardrails. When an agent encounters a task outside its expertise, it delegates the task to other agents – similar to professionals collaborating in the real world. Meanwhile, guardrails act as quality checkpoints that ensure that each agent validates the input or output before acting on it. These mechanisms let multiple agents operate somewhat independently, only interacting at defined handoff points.
Building AI agents with OpenAI SDK is more straightforward than others. It requires minimal boilerplate code and a manageable learning curve, using basic Python instead of framework-specific complex abstractions. The result is a solution that can scale easily.
Although developed by OpenAI, the SDK is not limited to their models and can be used with any LLMs. Additionally, built-in tracing makes debugging workflows simple, enabling users to pinpoint issues with precision.
Pydantic AI
Pydantic AI is an open-source multi-agent framework that allows for agentic interdependability, i.e., allowing one agent to invoke another in a subtask. The system leverages its own Pydantic library, enforcing strict data validation and type safety for agent interactions, which can reduce runtime errors caused by ambiguous or incorrect data.
Moreover, for more complex workflows, Pydantic AI offers a feature called Pydantic Graph, which lets you define multi-agent data flows as a directed graph of nodes and dependencies.
It provides highly readable code for developers where the structure and data you’re working with are clearly defined. Although it requires some upfront setup, the groundwork makes the code easier to maintain and a suitable investment for complex projects.
Pydantic AI also works with any LLM, and it can easily integrate with existing Python systems. Its lightweight Pydantic data validation, combined with support for continuous data streaming, makes it suitable for production systems and generally easy to manage and scale.
London • June 16 & 17, 2025
Speakers Gergely Orosz, Camille Fournier and Lara Hogan confirmed.
Takeaways
Going beyond single agents unlocks new capabilities. Whether it’s a crew of agents designing a product, a graph of agents managing a business process, or a duo of agents checking each other’s work, multi-agent frameworks enable AI to tackle tasks that are bigger, more complex, or more interdisciplinary than ever before.
By choosing the right framework for the job, developers can harness this power effectively. The landscape will surely keep evolving, but the concept of AI agents cooperating, much like humans do, is here to stay, opening the door to more ambitious AI-driven solutions in the years ahead.
