On September 10, 2026, Krea announced Krea Agents — an agentic platform that replaces manual ComfyUI graph wiring: the user describes the task with a single text prompt, and the agent assembles and launches a multi-stage generative pipeline from a pool of 60+ models. The system is available in an invite-only beta.

image
image

What happened

Krea published an announcement post and a launch video. The agent reads the canvas, plans a multi-stage pipeline, selects models from a pool of 60+ (Flux, Veo 3.1, Kling, Runway, Luma, Ideogram, and its own Krea 2), connects nodes, and launches the task. Krea 2 is a 12-billion-parameter model, a single-stream Diffusion Transformer with GQA blocks, QK-Norm, and SwiGLU-MLP. Planning and reasoning are performed by frontier LLMs, including Anthropic models, while generation runs on Krea and third-party models. The platform includes a persistent file system, the Context Library, with three levels: Session files, My Context, and Workspace Context, as well as skills — recorded step-by-step workflows that the agent repeats, OAuth connectors (Slack, Notion, Figma, Pinterest, Google Drive, YouTube, GitHub), and a host MCP server for agent-to-agent integrations. The beta is invite-only; applications are accepted at krea.ai/agents-beta.

Context

In creative production, multi-model generative pipelines have long been assembled manually: the user connected nodes in ComfyUI, selected models, and configured parameters, and each new task usually required rebuilding the graph. Meanwhile, the market is moving toward unified interfaces that combine dozens of generative models, and toward agentic assistants moving from chat into tools that perform tasks: the release mentions existing agentic stacks such as Claude Code, Cursor, and Codex. The Krea platform already unifies 60+ models under a single interface, so taking over pipeline orchestration is the next logical step for the company. Its own Krea 2 is built from standard architectural blocks typical of modern Diffusion Transformer models.

Why this matters for the industry

Krea moves the orchestration of generative pipelines from the user's hands to an agent with persistent file memory and reusable skills, lowering the barrier to entry for production pipelines: advertising campaigns, video, design. The aggregator of 60+ models becomes a single interface with payment via platform credits instead of separate API keys, and the host MCP server and OAuth connectors allow Krea to be embedded as a tool in existing agentic stacks rather than competing with them. If the reliability of the agent's plans is confirmed in practice, agentic orchestration with persistent context and skills could become a standard layer of production pipelines, and competition could shift from individual models to the quality of orchestration and accumulated proprietary skills and brand context. However, structural dependence on the platform — credits, cloud, and third-party frontier LLMs for planning — remains a cost and lock-in factor, which the host MCP server only partially mitigates.

Why this matters for users

For developers and creators, the main thing in this week's release is the invite-only beta window: you can apply at krea.ai/agents-beta while free LLM calls and additional generations are available this week. Instead of studying ComfyUI graphs, it is enough to describe the task with a prompt: brand context, mood boards (including those imported from Pinterest), and logos are placed in Workspace Context once, and the agent loads them itself thereafter. Repetitive workflows can be recorded as skills that the agent will reproduce step by step. This is an opportunity to test the "prompt instead of graph" scenario for free on real tasks and compare the quality of agentic planning with a manually configured graph.

What is still unknown / limitations

The announcement contains no paper, benchmarks, ablations, or reproducibility protocol, so the claim that a single prompt replaces manual ComfyUI graph wiring remains an unverified vendor marketing claim. The advantage of Krea 2 over open models of the same class is not shown or compared. Planning and reasoning are delegated to third-party frontier LLMs, meaning the "intelligence" of the system is an integration of other people's models, not Krea's own scientific achievement. The product is in an invite-only beta: there is no mass access, published pricing, or API yet, and the reliability of agentic orchestration is confirmed only by practice.

Sources

Author

Look at AI, editorial team