.svg "vast.png")
In my previous posts, I explored the importance of tools in developing AI Agents and discussed how agents can extend the capabilities of LLMs by using external tools....
Agent Orchestration Frameworks
I've covered What is an AI Agent and The Importance of Tools in Developing AI Agents in recent posts. In the next few posts I'll dive deeper into how AI Agents reason, plan and execute and several open-source frameworks popular for orchestrating AI Agents. I've covered a few of these frameworks at a high-level in recent posts (CrewAI and Huggingface Smolagents) and I'm looking forward to getting into the weeds.
HubSpot has a re-embed program for the engineering leadership team, the goal is to go back to being an IC for a period of 2-4 weeks and spend heads down time writing code with a specific team. For the next 3 weeks, I'll be re-embedding with our AI Agents and Skills teams and I'm very excited to dive into Agent Orchestration and learn from our extremely talented Software and Machine Learning Engineers. In preparation for this, I've been doing some pre-reading on open-source Agent Orchestration frameworks which the team have been exploring.
Orchestration Frameworks
As AI-powered agents take on more complex, multi-step tasks, a single model often isn’t enough. Orchestration frameworks help AI agents collaborate, plan, and execute tasks efficiently. Whether that’s automating workflows, coordinating specialised agents, or optimising problem-solving.
There are several cutting-edge approaches to orchestrating AI agents, each with its strengths:
• CrewAI – Assigns agents specific roles (like Planner, Writer, Reviewer) to function as a team. Great for business automation and structured workflows. You can see this in action with the rugby pundit crew.
• AutoGen – Microsoft’s framework for multi-agent conversations and distributed decision-making.
• Smolagents (CodeAgent) – A lightweight framework where AI writes and executes Python code to solve problems efficiently. Simple yet powerful for automating tasks.
• Plan Like a Graph (PLaG) – A prompting technique that improves how a single LLM plans by organizing tasks into a graph structure, making its reasoning more efficient.
Each approach has its ideal use case, some work best for automating business processes, while others shine in AI research, software development, or dynamic problem-solving.
CrewAI – Multi-Agent Teams with Defined Roles
Orchestration Method: Role-based collaboration (Planner, Researcher, Editor, etc.)CrewAI enables AI agents to work like a team, each with a defined role and a structured task pipeline. Inspired by real-world workflows, it’s ideal for breaking down complex tasks into manageable steps. For example, a blog-writing CrewAI might have:
- A Researcher gathering facts
- A Writer drafting content
- An Editor refining output
AutoGen – Conversational Multi-Agent Reasoning
Orchestration Method: Event-driven agent conversations (ReAct & self-improving loops)Microsoft’s AutoGen specializes in multi-agent conversations, where agents communicate, iterate, and improve their results. Agents can engage in structured debates, verification loops, and tool use, refining their responses like a team of experts.
For instance, a coding assistant might have:
- A Programmer agent writing code
- A Tester agent executing and debugging
- A Reviewer agent ensuring best practices
AutoGen’s event-driven model makes it highly scalable, suitable for large projects that require iterative refinement, long-term memory, and collaboration.
Smolagents (CodeAgent) – AI That Writes and Executes Code
Orchestration Method: ReAct (Reason + Act) via Python code execution
Unlike traditional AI agents that rely solely on reasoning, CodeAgent writes and runs Python scripts to solve problems autonomously. This makes it faster and more reliable for computational tasks. SmolAgents also have a smol course on aligning LLMs for your use case, one of our Senior MLEs recommended this course to me as mandatory reading, I'm a couple of notebooks in and have to strongly agree with them!
For example, given:
“Find the most downloaded Hugging Face model for text-to-video and return the details.”
A CodeAgent won’t just search for it, it will:
1. Write a Python script to call the API
2. Parse the response to extract the best result
3. Return the final answer
By integrating direct code execution, smolagents offers a lightweight, efficient solution for automating research, DevOps, and AI-powered workflows.
Plan Like a Graph (PLaG) – Smarter Planning for Single Agents
Orchestration Method: Graph-based prompting for structured planning
Unlike full orchestration frameworks, PLaG is a prompting technique that improves an AI’s ability to plan multi-step tasks efficiently. Instead of listing steps linearly, PLaG prompts an LLM to structure a plan as a graph allowing it to identify parallel tasks and dependencies.
For example, if an AI is planning a wedding, PLaG helps it determine that:
• Booking the venue must come before decorations
• Sending invitations can happen in parallel with food catering
• Final dress fitting must happen last
By enabling AI agents to think more like human planners, PLaG makes LLM-based systems faster, more structured, and more efficient. PLaG has an added benefit of easier testing and evaluation than alternative orchestration frameworks making it potentially a more attractive option for enterprise use-cases where reliability and reproducibility are key.
Conclusion
We're in a fortunate world at the moment where there are many open-source agent frameworks with a variety of strengths. Over the coming weeks I'm looking forward to spending dedicated time with each of these frameworks and sharing some more technical posts on each.