Comfy Agent: AI becomes the natural interface of the workflow
6 octobre 2026 · Kris Wilhem
On 1 October, ComfyUI presented Comfy Agent, an assistant able to intervene directly in workflows.
It can start from an existing workflow, change certain parameters, add and connect nodes, check for errors, run the graph, observe the results and start again if something does not work.
At first sight, it looks like one more assistant.
In reality, it is probably far more important than that.
A direction we were already exploring
At Axiris, this evolution interests us particularly because it overlaps with a path we had started to develop with our #comfyui channel (an IRC-style chat) with a connected LLM and a knowledge base dedicated to the Comfy documentation.
The workflow becomes steerable by conversation
Until now, working in ComfyUI meant understanding the logic of nodes, knowing the models, building the connections and knowing how to diagnose errors.
With Comfy Agent, a new layer appears: the conversation gradually becomes an interface for steering the workflow.
The user describes their intent.
The agent acts in the canvas.
The workflow can be modified, run and corrected without each action necessarily being carried out by hand.
The idea was already to have a chat interface linked to the documentation and to our knowledge base, so that the AI could better understand the ComfyUI environment, find the useful information and assist the user around their workflows.
That part is still at the draft stage with us, but the arrival of Comfy Agent clearly confirms that this direction is the right one.
In the cloud today, local tomorrow
Comfy Agent is currently offered in Comfy Cloud, with an arrival announced on Comfy Desktop.
The system currently relies on frontier models from Anthropic, but Comfy says it wants to open up the choice of LLM and also puts local models on its roadmap.
That is probably the point we will watch most closely.
The day this agentic layer can be cleanly connected to a local LLM, it will become possible to imagine an environment where:
- the workflow runs locally;
- the language model runs locally;
- the documentation is reachable through a knowledge base;
- the image and video models are part of the same system;
- the agent orchestrates the whole.
Value is shifting
If an AI becomes able to place the nodes itself, the value of the workflow does not disappear.
It shifts to a higher level.
The difference will increasingly be made on:
- the architecture of the system;
- the models available;
- the knowledge the agent has access to;
- the business rules;
- the orchestration choices.
Knowing how to connect a few nodes will remain useful.
But knowing how to build an environment able to understand why, when and how to use those nodes will become far more valuable.
What this changes for ComfyUI
Comfy Agent therefore marks an important change: the node canvas is no longer necessarily the only interface.
The conversation can become a layer above the graph.
And when that layer can truly talk to local models and private knowledge bases, ComfyUI will be able to evolve from an extremely powerful node tool into a genuinely agentic production platform.
That is the transition we are going to follow very closely.