DeepSeek Code: The Ultimate Form of AI Programming — From Chatbox to Terminal
📖 Glossary
AI Box (also known as Agent Computer / Agent PC), is a dedicated local hardware device that runs AI Agents. Pre-installed with an AI agent management system, plug-and-play, running 24/7. Users can remotely command AI to work via Discord, Slack, Telegram, WhatsApp, and more.
TL;DR: DeepSeek's funding surpasses 70 billion RMB, all-in on AI coding. DeepSeek-TUI, a terminal-based coding agent, hits 8,700 GitHub stars, working directly in engineering environments. When AI programming evolves from "Q&A" to "on-site collaborator," the terminal becomes its home turf. The Harness architecture benchmarks Claude Code, and the ultimate form of coding agents is taking shape.
1. 70 Billion Rounds of Ammunition, All In on AI Coding
In May 2026, DeepSeek confirmed total funding exceeding 70 billion RMB, against a backdrop of AI coding tools market growing at 45% annually, projected to exceed $50 billion by 2028.
More importantly, leadership declared: full focus on breakthrough AI research, abandoning short-term commercialization. This means DeepSeek won't chase "quick money" products under external pressure, but will concentrate on the hardest yet most valuable track in AI coding.
The first target: AI programming. DeepSeek is building a new Harness team, hiring Agent Harness PMs and engineers, with job descriptions explicitly stating "benchmarked against Anthropic's Claude Code."
2. DeepSeek-TUI: The Terminal Coding Agent
DeepSeek-TUI is the vanguard of this shift. Built in Rust by developer Hunter Bown, this terminal-based coding agent has earned 8,700 GitHub stars, becoming one of the most watched AI coding tools in the open-source community.
The fundamental difference from traditional chat-based AI coding tools (like ChatGPT Code Interpreter or Tongyi Lingma Web) lies in interaction boundaries:
- Chat AI: Like a remote code consultant — you send a message, it returns code snippets, you copy-paste to your editor. The flow is "you → AI → you → editor," with information shuttling between multiple interfaces
- Terminal Agent: Like an on-site collaborator — AI directly reads project files, understands code context, executes commands, runs tests, commits to Git. The flow is "you → AI → engineering environment," with AI directly operating your codebase
This difference seems simple but has profound implications. When AI can directly operate in engineering environments, the programming workflow fundamentally changes: you are no longer "the person writing code" but "the person directing AI to write code." Your role upgrades from "executor" to "architect + reviewer."
3. DeepSeek-TUI Core Capabilities
Based on project documentation and community feedback, DeepSeek-TUI's core capabilities include:
- Optimized for DeepSeek V4: Fully leveraging DeepSeek V4's long-context capability (128K+) to load entire project codebases at once
- Terminal-native interaction: No browser, no IDE plugin required — conversational programming directly in the terminal
- Visible reasoning process: DeepSeek's Chain-of-Thought reasoning displayed in real-time, showing you what AI is "thinking"
- File editing, shell execution, task management, sub-agent coordination: One agent handling multiple tasks, you just review results
- Fully local, extremely low cost: No internet, no API key needed, model runs locally, privacy and security fully controlled
Compared to Cursor ($20/month subscription), GitHub Copilot ($10/month), DeepSeek-TUI's local free model is extremely attractive for budget-conscious individual developers and small teams.
4. The Harness Architecture: Engineering Foundation for Coding Agents
The Harness architecture is the critical engineering framework in today's agent space, named after "horse harness" — equipping bare models with engineering "gear" to systematically address shortcomings in memory, code execution, and tool calling.
The technical progression of coding agents can be divided into four stages:
- Bare model stage: Direct code generation from LLMs (most primitive, represented by early ChatGPT Code Interpreter). Lacks engineering context, unstable code quality
- IDE plugin stage: AI assistance embedded in editors (represented by Cursor, GitHub Copilot). Has editor context, but AI is still "passive response"
- Terminal agent stage: AI directly operates in engineering environments (represented by Claude Code, DeepSeek-TUI). AI proactively executes, but requires human supervision
- Full Harness stage: Agent runtime with memory management, toolchain integration, execution environment isolation, multi-agent coordination (ultimate form, being built by DeepSeek Harness team)
The leap from stage 3→4 is fundamentally evolving from "one smart terminal assistant" to "a complete AI programming team."
5. Terminal Agents Need Terminal Hardware
The 24/7 operation requirement of terminal coding agents differs fundamentally from chat-based AI.
Chat AI works on a "you ask → it answers" model — if you don't ask, it doesn't work, so continuous operation isn't needed. But terminal agents work on a "you assign tasks → it autonomously executes → notifies you when done" model, where:
- AI needs to continuously monitor codebase changes (filesystem watcher)
- AI needs to continuously run tests, compilations, and static analysis (background processes)
- AI needs to keep working after you leave your desk (e.g., running an overnight test suite)
- AI needs to maintain project memory (what changed before, why it was changed)
This means you need dedicated hardware independent of your development machine to run coding agents continuously — it doesn't need a high-end GPU, but needs stability, low power, and 24/7 operation.
This is exactly Kaihe AI Box's use case: low-power 24/7 operation, physically isolated from your main development machine, giving agents a stable work environment. Your main PC handles gaming, video rendering, and daily work; Kaihe AI Box runs agents — both running without interference.
6. Prediction: The Ultimate Form of AI Programming
From DeepSeek-TUI's 8,700 stars, DeepSeek building the Harness team, and Anthropic's Claude Code iterating continuously, we can predict the ultimate form of AI programming:
You have a 24/7 online programming team, with members including: - Architect Agent (understands requirements, designs systems) - Coder Agent (writes code, fixes bugs) - Tester Agent (writes unit tests, runs integration tests) - Reviewer Agent (Code Review, security checks) - Deployer Agent (CI/CD, containerization, deployment)
These agents run on dedicated hardware, interact with you through a terminal, and are on call 24/7. You are no longer a "programmer" but a "programming team manager."
DeepSeek is building this future.
KaiheAiBox · AI Agent Tracker