An agent harness built for tasks that outlast your coffee break

What it does

DeerFlow is a long-horizon agent harness that coordinates sub-agents, memory, sandboxes, and extensible skills to tackle tasks lasting minutes to hours. It is a ground-up rewrite from ByteDance that explicitly targets deep research, coding, and creative workflows. The system runs as a Python backend with a Node.js frontend and integrates with a wide range of model providers.

The interesting bit

The project treats sandboxing and resource sizing as first-class concerns rather than afterthoughts. It ships with explicit deployment sizing guidance—starting at 4 vCPU and 8 GB RAM for local evaluation—and warns that improper deployment introduces genuine security risks, which suggests the authors have actually seen what happens when agentic code runs unchecked.

Key highlights

• Ground-up rewrite (v2 shares no code with v1) with active development moved entirely to the 2.0 branch
• Built-in support for sandboxed execution, long-term memory, and sub-agent delegation
• Integrates with Claude Code, Codex CLI, OpenRouter, vLLM, and standard OpenAI-compatible APIs
• Includes ByteDance’s own InfoQuest search and crawling toolset for research tasks
• Explicit security notice about deployment risks; Linux plus Docker is the recommended persistent server target

Caveats

• Version 2.0 is a complete rewrite, so expect early-adopter friction and breaking changes from any v1 experience
• Resource requirements are substantial: even light local evaluation wants 4 vCPU and 8 GB RAM, while shared server use starts at 8 vCPU and 16 GB RAM
• macOS and Windows are explicitly relegated to development or evaluation environments; production deployment is Linux-only

Verdict

Worth a look if you are building complex, multi-step agent workflows that need to survive for hours without human babysitting. Skip it if you just need a quick LLM wrapper or if your infrastructure budget tops out at a single Raspberry Pi.