Where Did the Once-Viral 'Crayfish' Go? Code Teardown Reveals It's Just a Subset of Claude Code

The multi-agent framework "Crayfish," which went viral half a year ago, seems to have faded from public view. But according to an analysis by a Bilibili creator, it hasn't disappeared—it has been absorbed into the development tools we use every day. Looking under the hood, Crayfish's most essential capability is essentially a subset of what Claude Code offers.

Crayfish's Core Capabilities: Decomposition and Execution

Crayfish's core functionality comes down to two things. First, it can break a complex problem into a combination of simpler problems, then assign an agent to solve each one. This decomposition strategy can significantly improve accuracy and capability in solving complex problems—the creator calls this Crayfish's "brain." He notes that back in the pre-LLM era, his team would manually decompose complex NLP problems and then combine multiple small models to improve success rates—the same underlying idea.

Second, Crayfish can operate various interfaces on a device to complete concrete tasks like web search, file deletion, and writing—these are its "hands and legs." The creator points out that these two most important capabilities have now been integrated into many development tools, whether it's their in-house TRIP, or Tencent's BOKBODY and CodeBODY—all contain similar features. In other words, Crayfish hasn't exited the stage; it has simply been absorbed into a larger tool ecosystem.

Multi-Agent Framework is a category of software infrastructure for organizing and orchestrating multiple agents to collaboratively complete tasks. It typically handles three things: task decomposition and assignment, inter-agent communication and collaboration, and aggregation and validation of execution results. Compared to single-agent systems, multi-agent frameworks use "divide and conquer" to reduce the processing complexity of individual models, thereby achieving higher success rates on complex tasks. Common implementation approaches in the industry include hierarchical (one master agent decomposes tasks and dispatches to sub-agents) and peer-to-peer (multiple agents negotiate as equals). The approach described in the article—"breaking complex problems into combinations of simpler problems, then distributing them to different agents"—is a typical form of hierarchical multi-agent framework.

Claude Code is a command-line coding agent tool for developers launched by Anthropic. It can read codebases, edit files, execute terminal commands, and invoke external tools—essentially a coding agent with "tool-calling" capabilities. When the article says Crayfish's capabilities are "equivalent to a subset of Claude Code," it means Claude Code also has built-in task decomposition, file operations, search, code generation, and execution capabilities, and is more complete in terms of engineering maturity. Claude Code can be understood as packaging the two capabilities of "decomposition + interface operation" into a mature development tool, which is why an independent multi-agent framework would appear to be a subset of it functionally.

From Skepticism to Company-Wide AI Development: A Team's Transformation

The creator shared his team's real experience. In September last year, when interviewing a R&D lead, he asked whether the candidate used AI to write code. The answer was no, with the reasoning that "AI code writing itself is too poor—it often produces irrelevant output and generates a lot of garbage code." This was the general perception of AI coding within the team in September and October of last year.

By January of this year, the situation had completely reversed—he had already required everyone to develop with AI. After the team adopted AI coding, productivity increased 3 to 4 times. The creator specifically clarified: some claim improvements of 150x or 100x, which he believes has a commercial component; 3 to 4 times is the real range that developers can actually perceive.

The key insight behind this shift is: whether the quality of tasks completed by AI is controllable still requires human judgment. The creator draws an analogy—when managing a 10-person team, the most demanding daily work is reviewing and checking team members' output; today, when developing with AI, the biggest problem is likewise decision-making and judging whether AI has completed tasks to quality standards, or whether it has produced large amounts of garbage or even mishandled things. You can add more AI to do reviews, checks, or even triple checks, but the final judgment still needs human oversight.

The productivity multiplier of AI coding needs to be understood with clear distinctions. A 3 to 4 times improvement typically refers to the net reduction in time spent on coding and debugging, under the premise that requirements are clear, the codebase is familiar, and testing and review provide a safety net. Claims like "100x" or "150x" often only count the instantaneous speed of code generation, ignoring the time spent on requirement clarification, result verification, and rework. When the creator emphasizes "the real range that developers can actually perceive," he is in effect reminding us that the benefits of AI coding depend on task type and verification cost—the more a step is bottlenecked by human review, the more the overall multiplier will regress toward the true value.

The Boundary of AI's Definition Is Always Moving Forward

The creator offers an interesting observation: there is a paradox in the history of AI development—when no one can solve something, whoever solves it has invented new artificial intelligence; but once a problem has been solved by humans, everyone feels "that doesn't count as AI anymore." So the boundary of AI's definition is always advancing.

He recalls that in 2025, capabilities demonstrated by small companies—long-term memory, personality cloning, automated web search—were all considered very AI; but this year or in another year or two, any agent or digital avatar can do these things, and they won't seem so "AI" anymore. He uses the power grid as an analogy: in industrial design, bearings and electricity—ordinary people don't need to deal with power plants; the national grid simply delivers electricity into homes. Crayfish is in that state today—it serves you every day, you just don't perceive it.

The creator also mentioned that throughout 2024 and 2025, while doing AI application development, it was already clear that many traditional software development and SaaS applications would be replaced by new AI tools within a year. Of course, traditional tools still have advantages, and many people are pivoting to follow new AI tools and offer similar services.

Permissions and Security: A New Solution to an Old Problem

When Crayfish went viral, there were also voices of opposition: give it permissions and it will control my computer. The creator believes these problems still exist today, but the solution depends on regulations and management approaches. Humans have always used tools within a social system—new tools bring new productivity and new value, but also new problems. He gives examples: after cars became faster, various safety regulations emerged; commercial aviation was also gradually refined. Today, discussing multi-agent systems similarly requires solving these issues within a year or two through internal rules or external regulations.

You may not have noticed, but the creator emphasizes that this is not just a matter for AI developers—it's the entire usage ecosystem: developers, users, and regulators all participating together. Crayfish represents a way of using multi-agent systems; it has already been integrated into work and tools. It hasn't disappeared—it has simply become part of tools and work, serving you every day, you just don't perceive it.