Goodbye AI Agents, Hello Agentic Workflows
Agentic Workflows are made up of smaller, lightweight AI Agents which acts as nodes in networks of agentic workflows…
I have written extensively about it before, the shift from AI Agents to Agentic Workflows.
In short, an Agentic Workflow is a workflow or mesh of nodes where each node is represented by an AI Agent.
But it is a more lightweight AI Agent than what traditionally was considered as an AI Agent.
And this pattern is forming the foundation of Agentic App building platforms like LangChain, OpenAI Agent Builder, Kore.ai AI for Process and more.
In fact, I would argue that LangChain is one of the first companies to have this vision.
A recent study from Microsoft took things a bit further, and I must say, it opened by eyes to what the future might look like…
Microsoft envisions Agentic organisation, where AI Agent nodes connect via a complex mesh (Agentic Workflows) solve complex problems by working collaboratively and concurrently.
Agentic workflows have the capability to split tasks and run processes in parallel, and bring it all together at the end.
In a recent article I shared a working Python notebook which walks you through the process.
So imagine if you have an augmented mesh of AI Agents within an organisation which work on tasks collaboratively.
Hence Microsoft introduced asynchronous thinking (AsyncThink) as a new paradigm of reasoning with large language models.
It organises the internal thinking process into concurrently executablestructures.
Specifically, Microsoft Research proposed a thinking protocol where an organiser dynamically assigns sub-queries to workers, merges intermediate knowledge and produces coherent solutions.
I think the organiser / classification agent is crucial in triaging and orchestrating the process.
Considering the OpenAI Agent UI prototype code GitHub, there are five agents, of which the triage agent directs the workflow. And acts as a fallback for the four other agents.
Experiments showed that AsyncThink achieves 28% lower inference latency compared to parallel thinking while improving accuracy on mathematical reasoning.
Scaling agentic organisation in two primary dimensions.
First is scaling the quantity of worker agents and how accuracy–latency trade-offs evolve as the agent pool capacity grows from a few to hundreds or even thousands.
Second is scaling the diversity of agents.
Agents could be fine-tuned for specific domains and be equipped with different external tools.
The human-AI agentic organisation is also a frontier is creating a Human-AI collaborative framework by integrating humans directly into the agentic organisation.
This could involve a human-as-organiser using the Fork protocol to dispatch tasks to AI workers, or a human-as-worker where the AI Forks tasks requiring human judgment.
Collaborative planning would allow humans and AI to co-design the asynchronous strategy before execution.
This path moves beyond pure AI autonomy to enable a powerful, hybridised intelligence.
Chief Evangelist @ Kore.ai | I’m passionate about exploring the intersection of AI and language. Language Models, AI Agents, Agentic Apps, Dev Frameworks & Data-Driven Tools shaping tomorrow.
