三足鼎立：读 Codex Harness、DeepSeek Harness 与 AgentScope 2.0 的横评

by Mycelium Protocol

---

原帖来源：社交媒体横评文章，对比三款开源 Agent 运行时底座。本文是读后感——对原文核心框架的二次整理，加上我们的补充判断。

---

一个扎心场景先说清楚

原文开头戳得很准。你做了一个 Agent，Demo 惊艳全场。一上生产，原形毕露：

• 任务跑到第 18 步，上下文爆了
• 昨天教过它的事，今天全忘了
• 用户一句恶意输入，它把 rm 命令直接怼到生产服务器上
• 半夜任务挂了，早上来一看：没有断点、没有日志、没有重试，一切从头再来

你以为该换个更聪明的模型？错。你缺的是 Harness——智能体运行时底座。模型只负责”想”，干活的是它底下那套工程系统。

上周 OpenAI 和 DeepSeek 一周内先后开源自家 Harness，加上阿里的 AgentScope，开源 Agent 底座已经三足鼎立。

---

Harness 的真实工作清单（苦活，不是玄学）

一个生产级 Harness 要兜住五件事：

能力	作用
任务循环	多步规划、失败重试、断点续跑
上下文管理	自动压缩、超大工具结果落盘只留占位符、防 token 爆炸
记忆持久化	跨会话记住用户偏好和任务进度，不是每次失忆重启
沙箱与审批	命令在隔离环境跑，危险操作必须人工点头
可观测	每个动作可追溯，出事了能回放取证

OpenAI 给过一组硬数据：同一个 GPT-5.6，裸奔跑分 13.3%，套上优化过的 Harness 直接拉到 38.3%，token 还省 6 倍。钱省在哪？就省在上下文压缩和推理保留这些工程细节上。

---

三强分析

选手一：Codex Harness（OpenAI）——出厂调校的整车

• 协议：Apache-2.0
• 规模：80+ Rust 子模块，9600+ 次提交，从 2025 年迭代至今，百万级用户生产验证
• 三个核心组件：
  • codex exec：面向 CI 的流水线执行器
  • Codex SDK（TypeScript/Python）：让开发者接入 Codex Agent 能力
  • app-server（JSON-RPC）：把 Agent 嵌进业务系统——持久化会话、流式事件、任务中断、自定义工具、人工审批；税务工具集成案例把处理时间砍了 1/3
• 核心定位：别把工作流硬塞进聊天框，把 AI 装进你的业务系统
• 短板：深度绑定 OpenAI 模型，数据出境到 OpenAI 云

选手二：DeepSeek Harness（dsh）——洞洞板组装底盘

• 协议：MIT（最宽松）
• 架构：Cordis 微内核，一切皆插件
• 四种运行模式：
  • 标准模式（全套工具）
  • 极简模式（仅 Bash + 编辑器）
  • PTC 模式（模型写 TypeScript 程序来完成多步操作）
  • Creation 模式（Agent 在运行时写/加载/卸载插件，自我进化）
• 模型无关：可以把 Claude Code、Codex 调度为子 Agent
• 对话日志：append-only，支持 fork/resume/replay
• 定位：给想自己搭底盘的团队用的原材料

选手三：AgentScope 2.0（阿里）——最完整的企业全家桶

• 架构：分布式部署，OpenTelemetry 埋点，对接 Higress/Nacos 的 MCP 生态
• 核心定位：国内企业私有化部署的最优解候选，Qwen 生态亲儿子，但也接 DeepSeek、OpenAI 兼容模型
• 多语言支持：Python / Java / TypeScript / Go（原来的 Workspace 抽象使 AGENTS.md 编辑等于升级 Agent）

---

三强速览对比

维度	Codex Harness	DeepSeek Harness	AgentScope 2.0
开箱即用	🥇	△	○
灵活度/换模型自由	△	🥇	○
私有化 + 多租户	△	○	🥇
协议宽松度	Apache	MIT 🥇	Apache

• 开箱即用：Codex Harness > AgentScope > DeepSeek Harness
• 灵活度：DeepSeek Harness > AgentScope > Codex Harness
• 私有化 + 多租户：AgentScope > DeepSeek Harness > Codex Harness
• 协议：DeepSeek（MIT）> AgentScope、Codex（Apache 系）

---

LangGraph 们还有价值吗？

原文的判断是：有价值，但战场换了。用分层视角看就清楚了：

• 传统框架（LangGraph / AutoGen / CrewAI）解决的是”多 Agent 怎么编排协作”——图结构、角色分配、消息路由。这个问题没消失
• Harness 解决的是”单个 Agent 怎么在生产里活下去”——上下文、记忆、沙箱、可观测性

两层不冲突，但优先级变了：没有稳定的 Harness 底座，LangGraph 编排得再漂亮也是沙上建塔。先把 Harness 选对，再谈编排层。

---

我们的补充：国内团队选型建议

优先考虑 AgentScope 2.0 的情况：

• 数据不能出境（医疗、金融、政务）
• 已在用 Qwen 系列模型
• 需要多租户隔离和 OpenTelemetry 接入现有监控体系

优先考虑 DeepSeek Harness 的情况：

• 想自己掌控底层，不接受黑盒
• 需要切换多家模型（DeepSeek / Claude / 本地 Qwen）
• 团队有 Cordis/插件生态经验，或愿意投入工程定制
• MIT 协议有商业授权优势

优先考虑 Codex Harness 的情况：

• 主力用 OpenAI 模型，不打算换
• 需要把 Agent 嵌进现有业务系统（app-server JSON-RPC 最省事）
• 看重百万用户生产验证和 9600+ 次提交的工程成熟度

一句话选型口诀：出境无所谓 + 用 OpenAI → Codex；不出境 + 用阿里云 → AgentScope；什么都想自己控 → DeepSeek Harness。

---

一句话总结

原文验证了这个时代的核心判断：Agent 能力瓶颈不在模型，在 Harness。三家开源底座各有侧重——Codex 是出厂整车、DeepSeek Harness 是原材料底盘、AgentScope 是企业全家桶。选哪个，看你的数据出境容忍度、模型绑定意愿和工程定制能力。

---

Mycelium Protocol — 追踪 AI 系统的底层演化

---

关于 Mycelium

菌丝协议。持续追踪 AI 工具、系统和实验的内容节点。

---

The Three-Way Harness Race: Codex, DeepSeek, and AgentScope 2.0

by Mycelium Protocol

---

Source: A social media comparison post reviewing three open-source agent runtime harnesses. This piece is a reading response — a second-pass synthesis of the original framework, plus our own supplementary analysis.

---

The Painful Production Scenario

The original post opens with something every agent builder recognizes:

Your agent demo was stunning. Then you shipped it to production:

• Task reached step 18 and the context window exploded
• Knowledge taught yesterday was completely forgotten today
• A single adversarial user input caused an rm command to fire on a production server
• A task hung overnight — no checkpoint, no logs, no retry. Start over from scratch

Think you need a smarter model? Wrong. What you need is a Harness — an agent runtime foundation. The model handles “thinking.” The engineering system underneath handles everything else.

Last week, OpenAI and DeepSeek each open-sourced their Harness within the same week. Add Alibaba’s AgentScope, and the open-source agent runtime landscape is now a three-way standoff.

---

The Harness Job List (Engineering Reality, Not Magic)

A production-grade Harness must handle five things:

Capability	Function
Task loop	Multi-step planning, failure retry, checkpoint resume
Context management	Auto-compression, oversized tool results stored to disk with placeholders, anti-token explosion
Memory persistence	Cross-session retention of user preferences and task state — not a fresh start every time
Sandbox & approval	Commands run in isolation; dangerous operations require human sign-off
Observability	Every action traceable; can replay events to investigate failures

OpenAI published hard numbers: the same GPT-5.6 scores 13.3% bare, 38.3% with an optimized Harness — and uses 6× fewer tokens. The savings come entirely from context compression and reasoning preservation engineering.

---

The Three Contenders

Codex Harness (OpenAI) — The Factory-Tuned Complete Car

• License: Apache-2.0
• Scale: 80+ Rust submodules, 9,600+ commits, iterating since 2025, million-user production verified
• Three core components:
  • codex exec: pipeline runner for CI
  • Codex SDK (TypeScript/Python): developer integration layer
  • app-server (JSON-RPC): embed agent in business systems — persistent conversations, streaming events, mid-task interruption, custom tools, human approval; a tax tool integration cut processing time by one-third
• Positioning: Don’t stuff workflows into a chat box — embed AI inside your business systems
• Weakness: Deep OpenAI model lock-in; all data exits to OpenAI cloud

DeepSeek Harness (dsh) — The Breadboard Chassis

• License: MIT (most permissive)
• Architecture: Cordis microkernel, everything-as-plugin
• Four run modes:
  • Standard (full tool suite)
  • Minimal (Bash + editor only)
  • PTC mode (model writes TypeScript programs for multi-step operations)
  • Creation mode (agent writes, loads, and unloads plugins at runtime — self-evolution)
• Model-agnostic: can schedule Claude Code and Codex as sub-agents
• Conversation logs: append-only, supports fork/resume/replay
• Positioning: raw material chassis for teams that want to build their own stack

AgentScope 2.0 (Alibaba) — The Complete Enterprise Suite

• Architecture: distributed deployment, OpenTelemetry instrumentation, MCP ecosystem via Higress/Nacos
• Positioning: the leading candidate for enterprise private deployment in China — native to the Qwen ecosystem, but also supports DeepSeek and OpenAI-compatible models
• Multi-language: Python / Java / TypeScript / Go (Workspace abstraction means editing AGENTS.md = upgrading the agent)

---

Head-to-Head Comparison

Dimension	Codex Harness	DeepSeek Harness	AgentScope 2.0
Out-of-box readiness	🥇	△	○
Model flexibility	△	🥇	○
Private deployment + multi-tenant	△	○	🥇
License permissiveness	Apache	MIT 🥇	Apache

• Out-of-box: Codex Harness > AgentScope > DeepSeek Harness
• Model flexibility: DeepSeek Harness > AgentScope > Codex Harness
• Private + multi-tenant: AgentScope > DeepSeek Harness > Codex Harness
• License: DeepSeek (MIT) > AgentScope, Codex (Apache family)

---

Are LangGraph and Friends Still Relevant?

The original post’s verdict: yes, but the battlefield shifted. A layered view makes it clear:

• Traditional frameworks (LangGraph / AutoGen / CrewAI) answer “how do multiple agents collaborate” — graph structures, role assignment, message routing. That problem hasn’t disappeared.
• Harness answers “how does a single agent survive in production” — context, memory, sandboxing, observability.

The two layers don’t conflict, but priority has shifted: without a stable Harness foundation, LangGraph orchestration on top is a house of cards. Choose the right Harness first, then worry about orchestration.

---

Our Supplement: Selection Guidance

Choose AgentScope 2.0 if:

• Data cannot leave the country (healthcare, finance, government)
• You’re already on the Qwen model family
• You need multi-tenant isolation and OpenTelemetry integration with existing monitoring

Choose DeepSeek Harness if:

• You want full control over the stack with no black boxes
• You need to switch between models (DeepSeek / Claude / local Qwen)
• Your team is comfortable with Cordis/plugin ecosystem investment
• MIT licensing matters for commercial use

Choose Codex Harness if:

• OpenAI is your primary model provider and you’re not switching
• You need to embed agents inside existing business systems (app-server JSON-RPC is the most turnkey path)
• You value million-user production verification and 9,600+ commits of engineering maturity

One-line decision rule: Data egress OK + using OpenAI → Codex. No data egress + Alibaba cloud → AgentScope. Want to control everything yourself → DeepSeek Harness.

---

One-Sentence Summary

The original post validates the central insight of this era: agent capability bottlenecks aren’t in the model — they’re in the Harness. The three open-source runtimes each have distinct positioning: Codex is the factory-tuned car, DeepSeek Harness is the raw chassis, AgentScope is the enterprise full-stack. Which to pick depends on your data egress tolerance, model lock-in appetite, and engineering customization capacity.

---

Mycelium Protocol — tracking the deep evolution of AI systems

© 2026 Mycelium Protocol. All rights reserved.