Dev & Engineering deep-researchmulti-agentlanggraphsandboxingsubagent-orchestrationtool-integrationim-channels

DeerFlow Super Agent Harness

Open-source long-horizon SuperAgent harness that orchestrates sub-agents, memory, and sandboxes to research, code, and create.

FollowAgents review · FARS-2.1
Use with care
63/ 100 5-point scale 3.2 / 5
1 2 3 4 5 6
1Trust17 / 29 · 2.9/5

Evidence shows: README details security considerations, including deployment risks and recommendations; supports sandbox mode, bash access control, file-write tools; provides user confirmation mechanisms (e.g., setup wizard); data flow transparency is evident in descriptions of agent behavior; sensitive data handling is documented (e.g., no password storage, HttpOnly cookies); dependency security has partial evidence (e.g., uv lock, CI tests), but no full audit or vulnerability scan; external effects are controlled (e.g., sandbox, permission limits); rollback mechanisms are mentioned (e.g., run cancellation, recovery); source attribution has partial evidence (e.g., copyright, contributor list), but no full supply chain verification. Deductions: insufficient dependency security evidence, incomplete source attribution.

2Reliability9 / 14 · 3.2/5

Evidence shows: README and code structure are consistent, providing multiple deployment methods (Docker, local dev) and configuration options; dependency availability is evidenced (e.g., uv management, mirror sources); failure messages are documented (e.g., error prompts, diagnostic tools). Deductions: no detailed error handling docs or troubleshooting guide.

3Adaptability12 / 18 · 3.3/5

Evidence shows: README targets multiple audiences (developers, researchers), provides multiple use cases (research, coding, creation); capability boundaries are described (e.g., sandbox mode, tool limits); trigger precision is described (e.g., multiple models, config options); environment fit is described (e.g., Docker, local dev, different OS). Deductions: no detailed scenario examples or boundary conditions.

4Convention12 / 18 · 3.3/5

Evidence shows: README is well-structured, providing install notes, config guide, examples; naming is stable (DeerFlow 2.0); known limitations are mentioned (e.g., no native Windows shell); license is MIT; versioning is mentioned (e.g., branch management), but no detailed changelog; maintenance responsibility is documented (e.g., contributing guide, security policy). Deductions: incomplete version changelog.

5Effectiveness9 / 13 · 3.5/5

Evidence shows: README describes output usability (e.g., report generation, file output); marginal value is described (e.g., integration with existing tools); cost-benefit is described (e.g., deployment sizing). Deductions: no actual performance data or user feedback.

6Verifiability4 / 8 · 2.5/5

Evidence shows: README claims are partially supported by code and docs; cross-source corroboration is limited (e.g., no third-party verification); fact-inference separation is documented (e.g., distinguishing official statements from community contributions). Deductions: insufficient cross-source verification.

Evidence confidence: Low Reviewed Aug 09, 2026 Reviewed revision e16ef2969b14
The upstream repository has new commits since this review. The score still applies to the reviewed revision shown and may not cover the latest changes.
Before you use it
  • Dependency security audit is insufficient; consider checking for vulnerabilities.
  • Source attribution is not fully verified; consider verifying supply chain.
  • Version changelog is incomplete; consider reviewing commit history.
Review evidence [1][2][3][4][5][6][7][8]
See the full review method →

What does this agent do, and when should you use it?

DeerFlow is an open-source super agent harness developed by ByteDance, built on LangGraph and LangChain, designed for long-horizon tasks that can take minutes to hours. It orchestrates sub-agents, maintains long-term memory, uses sandboxes for safe execution, and extends capabilities through a skill system. DeerFlow 2.0 is a ground-up rewrite that provides a full-stack solution including a web UI, Gateway API, and backend runtime, with Docker and local development support. Key features include skills & tools, session goals, context compaction, sub-agents, sandbox & filesystem, context engineering, and long-term memory. It integrates with observability tools like LangSmith, Langfuse, and Monocle, and can receive tasks from IM channels such as Telegram, Slack, Feishu, and more.

DeerFlow is a complete agent runtime that reads configuration (config.yaml) and skill definitions, calls multiple LLM providers (OpenAI, Anthropic, OpenRouter, vLLM, etc.) to execute tasks. It manages sessions and threads via the Gateway API, orchestrates sub-agents through LangGraph workflows, and executes code and file operations in local, Docker, or Kubernetes sandboxes. It provides built-in tools like web search, web fetch, file operations, and bash execution, and supports extending tools via MCP servers. DeerFlow can generate research reports, slides, web pages, images, and videos, and can receive user instructions through IM channels. It also offers CLI commands (make setup, make dev, make up) and a Python client (DeerFlowClient) for embedding.

  1. Researchers use DeerFlow for deep research, automatically collecting information and generating comprehensive reports with citations.
  2. Developers leverage DeerFlow's coding capabilities to generate code, tests, and documentation automatically in a sandbox.
  3. Content creators use DeerFlow to automate slide decks, web pages, and social media content generation.
  4. Enterprise teams interact with DeerFlow through IM channels (e.g., Feishu, WeChat) for task automation and information retrieval.
  5. Operations personnel use DeerFlow for data pipelines and dashboard generation and other automated ops tasks.
  6. Developers extend agent capabilities by integrating custom skills and MCP tools into their workflows.

What are this agent's strengths and limitations?

Pros
  • Provides a complete agent harness including frontend, gateway, and backend, ready to use with Docker deployment.
  • Highly extensible: customize tools via skills, MCP servers, and Python functions to adapt to various tasks.
  • Supports multiple LLM providers (OpenAI, Anthropic, OpenRouter, vLLM) including local models, avoiding vendor lock-in.
  • Built-in sandbox execution (local, Docker, Kubernetes) provides secure, isolated execution environments.
  • Integrates multiple observability tools (LangSmith, Langfuse, Monocle) for debugging and monitoring.
  • Mobile interaction via IM channels, suitable for various work scenarios.
Limitations
  • Deployment configuration is complex, requiring familiarity with Docker, Node.js, Python, etc.
  • High resource consumption; recommends at least 4 vCPU, 8 GB RAM, with higher requirements for production.
  • Version 2.0 is a complete rewrite, incompatible with 1.x, requiring migration effort for existing code.
  • Depends on third-party services and APIs, such as web search and LLM providers, which may incur costs.
  • Sandbox execution introduces security risks; improper configuration may expose systems.
  • Documentation and community support are still evolving, and some features may not be stable yet.

How do you install or deploy this agent?

  1. Clone the repository: git clone https://github.com/bytedance/deer-flow.git && cd deer-flow. 2. Run make setup to launch an interactive wizard to configure LLM provider, web search, and sandbox mode, generating config.yaml and .env. 3. For Docker development, run make docker-init to pull the sandbox image, then make docker-start to start services. For local development, run make check to verify prerequisites, make install to install dependencies, and make dev to start. Access at http://localhost:2026.

How do you use this agent?

After installation, interact with DeerFlow via the Web UI or IM channels. In the Web UI, create a new conversation, describe your task, and DeerFlow will autonomously plan and execute. Use slash commands (e.g., /new, /status, /memory) to manage sessions. For developers, use the Python client: from deerflow.client import DeerFlowClient; client = DeerFlowClient(config_path='config.yaml'); for event in client.stream('your task'): ... . You can also call the Gateway's LangGraph-compatible API for programmatic access.

FAQ

Is DeerFlow free?
DeerFlow itself is open-source under the MIT license and completely free. However, using it may involve costs for LLM API calls and server resources for deployment.
What API keys do I need?
At minimum, you need an API key for at least one LLM provider (e.g., OpenAI, Anthropic), and optionally a web search API key (e.g., Tavily). The setup wizard will guide you through configuration.
How long can DeerFlow handle tasks?
DeerFlow is designed for long-horizon tasks, using sub-agents, context compaction, and long-term memory mechanisms to handle tasks that take hours, but the actual duration depends on the model's context window and task complexity.
Which IM platforms does DeerFlow support?
Currently supports Telegram, Slack, Feishu/Lark, WeChat, WeCom, DingTalk, and Buzz. Once configured, it can automatically receive messages and process tasks.
How is sandbox security ensured?
DeerFlow provides Docker and Kubernetes sandbox isolation, but users must configure appropriate policies and follow security guidelines (e.g., not exposing unnecessary ports, controlling file access).

Compare agents like this one

The same FARS review applied across the shortlist this agent qualifies for.

Related agents