Skip to content
Back/Harness Engineering

Lightweight Harness vs Orchestration Harness

View in Graph
Updated 2026-08-31
2 min read
416 words

Lightweight Harness vs Orchestration Harness

Framework by @PawelHuryn: Harrison Chase is right that agents need memory (markdown, git, open formats), but personal agents and enterprise workflows require fundamentally different harness designs.

Blast-radius difference

Dimension Personal agent Business-critical workflow
Examples Coding copilots, email assistants Production pipelines, financial ops
Error impact Low — fix and move on High — revenue, compliance, safety
Harness style Lightweight Orchestration

Enforcement hierarchy

Security boundaries, file permissions, and tool access must be enforced in code, not merely "suggested" in prompts.

  • Lightweight harness solves portability and fast iteration.
  • Orchestration harness is where enterprise value is actually captured because it provides deterministic guardrails.

Key distinction

  • Lightweight harness: minimal routing, open formats, easy to move across models and environments.
  • Orchestration harness: robust state machines, strict access control, audit logging, failure-domain isolation, and human-in-the-loop escalation.

Practical recommendations

  1. Segment your projects: label each agent workflow as "personal" or "enterprise" and choose the harness depth accordingly.
  2. Move security down the stack: convert prompt-level "please do not delete files" into filesystem-level read-only mounts and code-level permission checks.
  3. Add deterministic frameworks for business-critical flows: explicit error handling, rollback procedures, and observability.

Counterpoints & Gaps

  • Many workflows start personal and later graduate to enterprise; harness migration is non-trivial.
  • "Enforced in code" assumes engineering resources that early-stage teams may not have.
  • There is no widely accepted standard or open-source reference for what constitutes a minimum viable orchestration harness.

Live contrast: Pi Agent vs DeepSeek Harness (2026-08-14)

来源:AI 简报 2026-08-14 Afternoon

@limbopeng 将 Pi Agent 与 DeepSeek Harness 作为两种相反路线的最新实例:

维度 Pi Agent(轻量核心) DeepSeek Harness(全层插件)
核心工具 read / write / edit / bash 四个 模型、工具、文件系统、Shell、沙箱、会话、Subagent、Loop 全部可替换
系统提示词 不到一千 token 无固定核心,官方 Coding Agent 只是默认组合
扩展方式 核心不变,复杂度推向外部扩展 没有不可替换的核心,所有层都是插件
控制权 交给外部扩展 Cordis 运行时统一调度

Pi Agent 是"先做减法再扩展"的极端:把 harness 压到最小,让外部工具和社区扩展承担多样性。DeepSeek Harness 是"先做加法再拆碎"的极端:先承认所有层都重要,再用可逆副作用和依赖重连让它们安全替换。两者不是简单的轻量 vs 重量,而是对"谁该拥有系统控制权"的不同回答。

这对个人和团队选择的启示是:如果你需要快速 setup 且愿意接受外部扩展的约束,Pi 路线更轻;如果你预计会频繁替换模型、工具链或执行环境,DSH 的可插拔抽象可能更持久。

Community reactions (2026-08-14)

来源:AI 简报 2026-08-14 Evening

  • Pi 作者被 DSH 启发:@mitsuhiko(Pi Agent 作者)表示 DeepSeek Harness 让他重新思考 Pi 的 harness 重构方向,说明两种路线正在相互渗透而非简单对立。
  • Token 成本对比:@BohuTANG 实测 DSH 默认挂载 25 个工具,token 和成本约为 Pi 的 3 倍;插件代码上的松耦合不等于模型行为也能模块化。这提示"可组合性"的收益需要以实际运行成本衡量。

2026-08-31 更新:扩展栈与极简主义的同场冲突

来源:AI Briefing 2026-08-31 Morning

2026-08-31 的时间线同时出现了 Pi Agent 扩展栈推荐和对 Oh My Pi 的批评。扩展栈方向强调网络访问、长时记忆、todo overlay、子代理、MCP 适配和 context compression。它把轻量核心保持不变,把复杂能力推到可选扩展。

反向观点认为,多个 benchmark 优化技巧捆绑在一起后会形成不可见维护负担。每次模型发布都可能改变技巧的收益。如果没有按模型版本重新评估每个组件,扩展栈可能让 Agent 管理自身脚手架而不是完成任务。

这补充了本页的核心区分:轻量 harness 的优势不是扩展少,而是默认控制面小。可选扩展需要按任务风险、当前模型和验证成本单独进入,而不是作为永久默认项。

Sources

Synthesized from 4 sources
  • AI 简报 2026-04-13Supporting source listed by this page.Whole pagemediumbody
  • AI 简报 2026-08-14 AfternoonSupporting source listed by this page.Whole pagemediumbody
  • AI 简报 2026-08-14 EveningSupporting source listed by this page.Whole pagemediumbody
  • AI Briefing 2026-08-31 MorningSupporting source listed by this page.Whole pagemediumbody

Evolution

1 event
  1. absorbed

    Derived from source material

    This page is currently synthesized from 4 sources.

    From AI 简报 2026-04-13, AI 简报 2026-08-14 Afternoon, AI 简报 2026-08-14 Evening, AI Briefing 2026-08-31 MorningTo Lightweight Harness vs Orchestration Harness
    Sources: raw/briefing/AI Briefing/2026-04-13.md · raw/briefing/AI Briefing/2026-08-14-11-31.md · raw/briefing/AI Briefing/2026-08-14-18-29.md · raw/briefing/AI Briefing/2026-08-31-08-11.md

Linked from