 General-purpose agent powered by DeepSeek Harness. Handle any task — coding, research, data analysis, writing, and automation — in a persistent sandbox with terminal, filesystem, browser, and live VNC desktop. ## What it's good at - **Coding** — write, run, debug, and refactor programs in any language; scaffold and maintain complex projects - **Research** — scrape the web, synthesize information from multiple sources, produce structured reports - **Data analysis** — process CSV, JSON, and Excel files; run analysis; generate charts and dashboards - **Writing** — draft reports, proposals, emails, documentation in Markdown or HTML - **Automation** — build scripts, CLI tools, and pipelines to automate repetitive workflows - **Live preview** — run web apps, scripts, and tools in the sandbox and see results in real time via VNC ## How it works DeepSeek Harness runs in a persistent Docker sandbox with a full desktop environment. It has terminal access, a browser, filesystem, and network — essentially a complete machine at your disposal. It uses DeepSeek V4 Pro or V4 Flash as its reasoning engine. You describe what you need, and it plans the approach, executes the work, observes the result, and iterates until it's right. The sandbox is `longrunning` — it pauses when idle and resumes when you return. Your code, files, and project state persist across conversations indefinitely. If Docker sandbox is not enabled, tasks execute directly on the machine where Yao Engine is installed. ## Best practices **Describe the outcome and constraints.** Tell it what you want to be true when it's done, plus any boundaries. Don't dictate how to get there — let it figure out the approach. "Find out which products are driving revenue growth and show me a breakdown by region" is enough. **Give it context upfront.** Attach files, paste screenshots, reference documents already in the workspace, or describe the situation in detail. The more context it has, the fewer back-and-forth rounds you'll need. **For complex tasks, break it into phases.** Ask it to research first, then plan, then execute. This avoids solving the wrong problem. Example: "First look at this spreadsheet and tell me what data is available. Then suggest what analysis would be useful." **Give it a way to verify.** If you can describe what "done" looks like — a file in a specific format, a chart that shows a trend, a report with specific sections — say so. It will check its own work before delivering. **Keep the conversation going.** Each message builds on everything before it. Say "now add a comparison to last year" or "make it shorter" — it keeps the full context and picks up right where you left off. **Examples:** > Research the top 5 competitors in [industry], summarize their pricing models, and produce a comparison table. > Analyze the attached sales-report.xlsx. Which products are growing fastest? Build a dashboard with the key metrics. > Draft a project proposal for [topic]. Audience: executive team. Keep it under 3 pages. Include a timeline and budget estimate. > Monitor this RSS feed daily and send me a summary of anything related to [topic]. Save results to a running log. ## Capabilities at a glance | Capability | Details | |------------|---------| | Model | DeepSeek V4 Pro / V4 Flash (configurable) | | Languages | Python, Node.js, Bash, and anything installable | | Web access | Browse, scrape, fetch APIs | | File I/O | Read and write any file format | | Preview | Live VNC desktop — see apps running in real time | | Sandbox | Persistent Docker desktop, full terminal and browser | | Lifecycle | Long-running — pauses on idle, resumes on return | | Workspace | Permanently stored — code and files persist indefinitely | ## What's next - [Claude Code](/docs/en-us/built-in-agents/claude-code) — same sandbox capabilities, powered by Claude - [Applet Workshop](/docs/en-us/built-in-agents/applet) — for visual mini-tools with instant preview - [AI Experts: Find the Right One](/docs/en-us/getting-started/ai-experts-intro) — see all built-in experts
DeepSeek Harness
feedback
反馈