Comfy Org has announced the start of beta testing for Comfy Agent, a built-in AI assistant in ComfyUI. The assistant builds and edits node-based workflows based on text prompts, runs generations, and fixes errors, becoming a ready-made copilot for working with the editor.

image
image
image

What happened

The launch was announced in the official ComfyUI stream "Comfy Agent Beta: VFX, Marketing & E-Commerce Workflows," which took place on September 10, 2026. According to the official product announcement, Comfy Agent works in both the local version of ComfyUI and Comfy Cloud, while the Agent Tools documentation clarifies that the in-app agent is available to users with an active Comfy Cloud subscription. Applications to participate in the beta are accepted via a Typeform form, and the product page has already been published at comfy.org/agent.

Context

ComfyUI already had a way to connect external AI agents: Comfy MCP, an open interface that connects the editor to agents like Claude and Cursor via the Model Context Protocol (MCP). Comfy Agent differs from this approach: instead of an open protocol for external tools, it is a ready-made out-of-the-box solution built into the editor itself. In the official Agent Tools documentation, Comfy MCP, Comfy In-App Agent, and Comfy CLI are described as tools operating at the same level, so the novelty of Comfy Agent lies in integration and packaging, not in a new model. Technically, the agent layer reduces to generating a node graph (directed acyclic graph) and tool use via the ComfyUI API. The configuration is hybrid: the agent backend is deployed in Comfy Cloud, while local models are claimed for simple tasks.

Why this matters for the industry

For the industry, this is a capture of the "agent" layer in content generation pipelines: the "text → working workflow" connection becomes a built-in capability of the editor, tied to a Comfy Cloud subscription. Since, according to the official documentation, using Comfy In-App Agent requires an active Comfy Cloud subscription, Comfy Org is turning the agent into a monetization point, pulling users onto its own cloud infrastructure. This squeezes the niche of startups that built their product as a wrapper for "ComfyUI + external agent via Comfy MCP," taking away their key differentiator before they can build an audience. The signal for product teams is clear: the automation layer for custom content generation workflows (VFX, marketing, e-commerce) is becoming native, and the agent is becoming a standard component of the pipeline.

Why this matters for users

For users working in ComfyUI, this means that right now they can apply for the beta via the form and get an assistant that builds complex node chains based on a simple text description, rather than requiring manual selection of each node. The official stream from September 10, 2026, allows you to see how the agent builds workflows for VFX, marketing, and e-commerce, reproduce these scenarios on your own tasks, and compare the agent's build speed with manual. However, it is still too early to connect the agent to a production pipeline: there is no SLA, public API, pricing, or latency measurements, so a reasonable step now is to collect your own quality and response measurements on your own scenarios.

What is still unknown / limitations

The system has no independent evaluation: there are no benchmarks, data on the frequency of success and failure, generation latency and their cost, and the entire evidence base is limited to a demonstration in the official stream. The agent architecture, model weights, and model provider have not been disclosed, so the system cannot be reproduced or independently verified. The methodology for finding compatibility errors has not been disclosed. There is no data on the proportion of tasks the agent completes without manual intervention and the frequency of errors.

Sources

Author

Look at AI, editorial team