DeerFlow: Super-agent framework from ByteDance

Contents

OpenClaw has shown what LLMs are capable of when agents use them correctly. However, the dangers associated with this are not only great but even unmanageable: In some cases, OpenClaw stole credit cards or took over the entire computer. Nevertheless, the technology holds enormous potential, which has found some imitators trying to correct the weaknesses.

ByteDance takes the lead

One of the first players to emerge from hiding is ByteDance, the parent company of TikTok, with the tool DeerFlow (Deep Exploration and Efficient Research Flow). After the project has already received over 56,000 stars on GitHub, it's worth taking a closer look.

ByteDance wants you to sign up for a coding subscription to use it, but it's not necessary: DeerFlow can also be used with local models. However, local use becomes a bit more difficult when the framework is supposed to perform web searches independently. This is extremely useful, as the knowledge cutoff of most LLMs is considerably in the past and they are not informed about new developments. Here you can add the search engine Tavily, for example, which has a free version that allows a thousand search queries per month. It is also available at InfoQuest, which ByteDance offers as proprietary. However, the number of queries is limited to a thousand per account in total, and prices for more are not directly available. Interestingly, ByteDance does not offer this service in China at all, possibly for censorship reasons.

All can be installed and configured locally. The detailed instructions can be found on GitHub. First, you need to adjust a configuration file in YAML format. Then, Docker containers can be started, in which the system then runs. The containers open a port (2026, is that a coincidence?) on the host system. The whole thing works quite unproblematically if you write all environment variables in .env, as the system does not use the shell variables. The result looks very similar to the DeerFlow playground, but it can be supplemented with new chats.

When playing with the playground, you can also tell the tool its origin in Deep Research. The agent can research complex topics itself and achieves good results, which also include new developments. The suggestions for heise-Verlag (see Figure 1) are, as usual with other agents, very extensive and sound sensible. However, that's far from all.

From Deep Research to Super Agents

Internally, DeerFlow is based on LangChain and LangGraph, making it excellently extensible. ByteDance also emphasizes this and calls the framework a "Super Agent Harness" – a framework that can be excellently extended. Reportedly, the ByteDance developers redesigned DeerFlow when they recognized this extensibility of LangChain as a crucial point.

Such extensions can occur at different levels. Central to this are skills and tools, in which users define workflows via Markdown files. The research workflow used in the example in the figure is also based on such a skill. DeerFlow loads skills provided by users on demand when they are needed. This saves memory and computing time and allows for dynamic extension of the system at runtime. Tools are somewhat more advanced and perform complex tasks such as web search, but can also safely execute Bash scripts or Python programs in a sandbox.

ByteDance has recognized that complex tasks are difficult to complete in a single pass. Therefore, the concept of sub-agents, which have their context, their tools, and completion conditions, originates from this. This allows them to run in parallel with each other, saving time.

The sandbox in DeerFlow works like its own computer. Each process receives its own environment with a separate file system, which in turn can contain skills, etc. This isolation allows the processes to act independently of each other.

ByteDance places great importance on the isolation of contexts. Each agent has its own context and is not disturbed by other agents. Incidentally, this also allows for the aforementioned parallelization of agents. To keep the context as small as possible (and thus save memory or tokens in the LLM), summarization takes place continuously.

DeerFlow has long-term memory and remembers content beyond a conversation. It thus functions as a knowledge store and adapts better and better to the users' wishes. It stores all this data locally. ByteDance is very familiar with recommendations and apps that users don't want to do without. However, unlike TikTok, DeerFlow creates real added value.

DeerFlow brings new ideas

With DeerFlow, ByteDance is releasing an agent framework with many new ideas and showing how agents could develop in the future. Particularly exciting are the easily implementable skills, the isolation and parallelization, and the sandboxes, which are intended to avoid dangerous situations like those with OpenClaw.

Agentic AI is currently occupying researchers very intensively. Nobody knows exactly how the field will unfold yet. Alternative tools like n8n or Dify are still much more static here than DeerFlow, but new directions may now emerge for them as well. It will be all the more exciting to observe how the DeerFlow framework continues to develop.

(dahe)