DeepSeek Harness vs Claude Code, Codex, WorkBuddy, Manus, and OpenClaw: What's Actually Different?
New agent names show up every week: Claude Code, Codex, WorkBuddy, Manus, OpenClaw. When DeepSeek Harness (DSH) landed, the obvious question followed — how is it different from all of them?
This post answers that from a more fundamental angle: these tools don't sit on the same layer. DSH is a composable, self-modifiable agent runtime. Most of the rest are finished products, closed or only semi-open.
The short version: different axes
- DeepSeek Harness is the framework itself. Tools, sandbox, approval flow, subagents — all assembled from Cordis plugin rows, and even the runtime hosting the agent is open to modification.
- Claude Code / Codex are coding products: one in your local terminal, one in a cloud sandbox. Both exist to write code, and the vendor draws the extension boundary.
- WorkBuddy / Manus are task products: one drives your desktop, one hands off general tasks to a cloud agent and returns the deliverable.
- OpenClaw is a personal assistant: self-hosted, always on, managing your life and devices over messaging channels.
One line: other products hand you an agent in a box. DSH hands you the agent in parts.
Six agents, side by side
| Axis | DeepSeek Harness | Claude Code | OpenAI Codex | WorkBuddy (Tencent) | Manus | OpenClaw |
|---|---|---|---|---|---|---|
| Shape | Programmable agent framework + Web GUI | Terminal CLI | Cloud coding agent | Desktop app | Cloud autonomous agent | Open-source assistant framework |
| Where it runs | Local Node process + browser UI | Local terminal | OpenAI cloud sandbox VM | Your desktop | Cloud VM (E2B) | Self-hosted gateway, always on |
| Main job | General agent building & orchestration | Writing code | Repo-scale code | Office automation, driving the OS | General research & delivery | Personal life tasks |
| Extension model | Cordis plugin rows, versioned + approved + rollback | Hooks, subagents, skills, MCP, plugins | Closed, product updates only | Visual task flows, low-code | Closed, plus an API platform | Plugins, skills, channel integrations |
| Security model | Local file sandbox + stepped approval escalation | Local permission prompts | Cloud isolation + git sync | Desktop permission control | Cloud sandbox isolation | Self-hosted, channel-level control |
| Open source | ✅ | ❌ (CLI free, models paid) | ❌ | ❌ | ❌ | ✅ |
Breaking each one down
DeepSeek Harness: the runtime you can rewrite
DSH's philosophy is "everything is a plugin row": file tools, sandbox, approval stack, model routing, the subagent registry — each is one line in cordis.yml. A session is assembled from a Host composition (persistent, cross-session services like storage and sandbox) and an Agent Preset (the tools, persona, and prompts mounted per session).
Three mechanisms set it apart:
- A dynamic plugin lifecycle: define → run → update → rollback → undefine temporary plugins mid-session. Packages are immutable and versioned, and client-side code needs your approval (single check = this version only, double check = authorize future versions). Nobody else offers that combination of in-runtime extension plus audit.
- A two-plane architecture: Host (the Node process handling files, network, commands, and model tools) and Client (the browser GUI handling themes, slot UI, and page state), with clean separation of responsibilities.
- Four orchestration primitives: long-running Goals in the same session, Subagents (forkable to inherit context), Workflows (scripted large-scale fan-out), and Ralph (fresh-context iteration loops).
The deeper point is framework reflexivity: the Harness source checkout lives on your machine, so you can read and edit your own presets and compositions. In other products the tool set is fixed by the vendor; in DSH it's a config file you own. GitHub repo | Architecture breakdown
Claude Code: closest cousin, different job
Anthropic's terminal coding tool runs locally, edits your files, and executes shell commands. Extension comes from hooks, subagents, skills, MCP, and plugins, plus the Agent SDK for developers.
The difference: Claude Code is a CLI product for coding tasks, and its extensions are hook-style (intercept events, add tools via MCP). DSH is a general agent runtime, and its extensions are compositional (capability = plugin row, session = preset assembly), with dynamic plugins, version rollback, and a Node + browser dual platform. Claude Code is great at writing code; DSH is built for assembling agents.
OpenAI Codex: the coding agent in a cloud sandbox
Codex runs tasks in parallel inside OpenAI-managed cloud VMs, syncs your repo through git (triggered by GitHub PRs and issues), and never executes directly on your machine. A parent task can spawn child tasks (delegation). It ships as a CLI, an IDE extension, and a web app. Codex handbook | Codex vs Claude Code
The difference: Codex is a closed cloud product centered on the repo, with the sandbox in the cloud and internals you can't touch. DSH runs in a local process, with sandbox policy you can see and configure and every tool and orchestration path open source.
WorkBuddy (Tencent): the visual desktop agent
Tencent's productivity agent is a desktop app built around "AI operating your computer": moving the mouse, driving the browser, organizing files. It bundles an AI-native knowledge base, uses visual task flows (Vibe Working) for orchestration, and offers team collaboration in the enterprise tier. Tencent Cloud techpedia | iFanr hands-on
The difference: WorkBuddy targets office users with low-code, point-and-click interaction, and its object is the GUI desktop. DSH targets developers, expresses everything in code and config, and its object is the agent runtime itself.
Manus: the hands-off general task agent
Manus focuses on autonomous general tasks: web search, browser control, running code in an Ubuntu sandbox, a VS Code environment, and internal multi-agent collaboration. It runs asynchronously and delivers a report, spreadsheet, or site when done, and it has opened an API platform in recent years. Architecture analysis | E2B sandbox mechanics | Heise review
The difference: Manus is a task finisher — give it a goal, it completes it in the cloud and returns a result, with the process a black box. DSH is a runtime — the process and every capability are transparent, auditable, and rollback-able.
OpenClaw (formerly Clawdbot/Moltbot): the open-source life assistant
A self-hosted, always-on assistant: one gateway running 24/7, plugged into WhatsApp, Telegram, Discord, and more, able to control computers and phones, automate a browser, take voice input, and keep cross-session memory — a personal Jarvis. It has been renamed Clawdbot → Moltbot → OpenClaw. DigitalOcean guide | Milvus explainer | O'Reilly review
The difference: OpenClaw's territory is personal life and messaging channels (replying, scheduling, controlling devices); DSH's territory is software development and agent building. One is bounded by channels and devices, the other by plugins and sessions.
How to choose, in one line each
- Want AI to do work for you (office tasks) → WorkBuddy / Manus
- Want AI to write code in your terminal → Claude Code
- Want AI to batch-edit repos in the cloud → Codex
- Want a personal assistant for life → OpenClaw
- Want an agent framework you can take apart, modify, assemble, approve, and roll back — one where the framework itself is the thing you build on → DeepSeek Harness
Wrapping up
The agent space is splitting in two: on one side, increasingly "ready out of the box" task products; on the other, increasingly programmable runtimes. DeepSeek Harness is clearly betting on the latter, treating composability, auditability, and rollback as first-class concerns. If you want to start from zero, the quick start is the fastest way in; if you want to go deeper on a specific plugin or practice, keep browsing Oh! dsh.