docs: sharpen cross-harness positioning

Author: wxj630

README.md

@@ -1,4 +1,4 @@
-# Bayesian-Agent
+# Bayesian-Agent: A Bayesian Self-Evolving Agent Framework with Cross-Harness Adaptation
 
 <div align="center">
   <img src="assets/banner.png" width="920" alt="Bayesian-Agent banner"/>
@@ -10,13 +10,19 @@
   <a href="https://github.com/DataArcTech/Bayesian-Agent">GitHub</a>
 </p>
 
-Bayesian-Agent is a standalone Bayesian self-evolving agent framework for turning agent failures into reusable, evidence-weighted Skills and SOPs.
+Bayesian-Agent is a Bayesian self-evolving layer for turning agent failures into reusable, evidence-weighted Skills and SOPs across agent frameworks and execution harnesses.
+
+It is designed to stand out from monolithic agent frameworks in three ways:
+
+- **Run from scratch**: start with no prior traces and evolve Skills during full benchmark or production runs.
+- **Repair incrementally**: attach to an existing agent, read its failed trajectories, and rerun only the tasks that need repair.
+- **Adapt across harnesses**: integrate with GenericAgent today, and with other agent frameworks through a portable trajectory schema and adapter boundary.
 
 > v0.4 is the first standalone release. It includes the core Bayesian Skill Evolution package, schemas, CLI utilities, experiment artifacts, and a clean GenericAgent integration boundary. GenericAgent itself is not copied, vendored, or forked.
 
 ## 📅 News
 
-- **2026-05-09:** Released Bayesian-Agent v0.4 as a standalone package with core Bayesian Skill Evolution primitives, schemas, CLI utilities, and experiment artifacts.
+- **2026-05-09:** Released Bayesian-Agent v0.4 as a standalone cross-harness Bayesian Skill Evolution package with schemas, CLI utilities, and experiment artifacts.
 - **2026-05-09:** Added the optional GenericAgent adapter boundary without copying or vendoring GenericAgent.
 - **2026-05-09:** Published bilingual project documentation and the Bayesian-Agent framework diagram.
 
@@ -38,12 +44,12 @@ In that setting, **Skills** and **SOPs** become first-class engineering assets.
 - which failure modes to avoid
 - when to stop, retry, or rewrite the procedure
 
-Bayesian-Agent asks a simple question: if Skills are hypotheses about how to solve tasks, why should they evolve by anecdote instead of evidence?
+Bayesian-Agent asks a simple question: if Skills are hypotheses about how to solve tasks, why should they evolve by anecdote instead of evidence? The answer is a framework-agnostic evolution layer that can bootstrap Skills from scratch, repair existing agents incrementally, and move across harnesses as long as they emit verified trajectories.
 
 <div align="center">
   <img src="assets/bayesian_agent_overview.png" width="900" alt="Bayesian-Agent overview"/>
   <br/>
-  <em>Bayesian-Agent turns verified agent trajectories into posterior-weighted Skills and SOPs.</em>
+  <em>Bayesian-Agent turns verified trajectories from any compatible harness into posterior-weighted Skills and SOPs.</em>
 </div>
 
 ## 🧠 Core Idea
@@ -83,8 +89,9 @@ After each verified trajectory, the framework updates a posterior belief over th
 - **Bayesian Skill registry**: maintain Beta posteriors, failure modes, token cost, latency, turns, and context distribution.
 - **Failure-mode-aware repair**: identify recurring errors and generate focused repair plans.
 - **Token-aware context building**: inject only the Skills with useful posterior evidence.
-- **Full and incremental modes**: evolve from scratch or attach to an existing agent as a repair layer.
-- **Framework-agnostic boundary**: integrate with GenericAgent and other harnesses through adapters instead of vendoring their code.
+- **Full self-evolution from scratch**: run all tasks, collect evidence online, and evolve Skills without prior traces.
+- **Incremental repair for existing agents**: consume failed trajectories from a baseline agent and rerun only the failed tasks.
+- **Cross-harness adaptation**: integrate with GenericAgent today and other agent frameworks through adapters instead of vendoring their code.
 - **Standard-library-first core**: v0.4 has no runtime dependency beyond Python.
 
 ## 🧬 Self-Evolution Mechanism
@@ -197,7 +204,7 @@ skill_context = SkillContextBuilder(registry).render(task_context="sop_bench")
 print(skill_context)
 ```
 
-## 🔁 Two Operating Modes
+## 🔁 Three Operating Patterns
 
 ### 🌱 Full Self-Evolving Mode
 
@@ -215,6 +222,16 @@ Base Agent -> Failure Traces -> Bayesian Skill Evolution -> Rerun Failures -> Hi
 
 This is the recommended production path because it improves an existing agent without retraining the model or replacing the original harness.
 
+### 🔌 Cross-Harness Adaptation Mode
+
+Bayesian-Agent is not tied to a single agent runtime. Any agent framework can become a backend if it emits the common trajectory schema and accepts posterior-weighted Skill context through an adapter.
+
+```text
+Any Agent Harness -> Trajectory Schema -> Bayesian Skill Registry -> Adapter -> Next Harness Run
+```
+
+This makes Bayesian-Agent a portable Skill/SOP evolution layer rather than another closed agent framework.
+
 ## 📊 Experimental Results
 
 The v0.4 prototype was validated with GenericAgent and `deepseek-v4-flash` on SOP-Bench and Lifelong AgentBench.
@@ -247,13 +264,13 @@ In incremental mode, Bayesian-Agent only reran failed GenericAgent tasks:
 | SOP-Bench | GA+BayesianIncremental | deepseek-v4-flash | 100% | 216k | 10k | 226k | 17.73 |
 | Lifelong AgentBench | GA+BayesianIncremental | deepseek-v4-flash | 100% | 71k | 7k | 78k | 25.57 |
 
-The result shows that Bayesian-Agent can work as a plug-in repair layer: it can take an existing agent below 100% accuracy and improve it with a small amount of incremental inference.
+The result shows that Bayesian-Agent can work as a plug-in repair layer: it can take an existing agent below 100% accuracy and improve it with a small amount of incremental inference. This is the practical advantage over one-off benchmark agents: Bayesian-Agent can sit beside a harness, learn from its failures, and improve it without replacing it.
 
 Experiment artifacts are stored under [`artifacts/`](artifacts/), and the method note is in [`docs/method.md`](docs/method.md).
 
-## 🔌 Relationship to GenericAgent
+## 🔌 GenericAgent and Cross-Harness Adaptation
 
-The first prototype was validated inside GenericAgent, but Bayesian-Agent is not a GenericAgent fork.
+The first prototype was validated inside GenericAgent, but Bayesian-Agent is not a GenericAgent fork and not just a GenericAgent add-on.
 
 The open-source structure is:
 
@@ -263,7 +280,9 @@ The open-source structure is:
 - `schemas/`: portable trajectory and Skill belief schemas
 - `artifacts/`: reproducible benchmark result files
 
-GenericAgent remains an optional backend. Users can integrate Bayesian-Agent with their own agent harness by emitting the common trajectory schema.
+GenericAgent remains the current experimental backend. Users can integrate Bayesian-Agent with their own agent harness by emitting the common trajectory schema and implementing the adapter boundary.
+
+The long-term direction is to make Bayesian-Agent the Bayesian Skill/SOP evolution layer for many agent runtimes: GenericAgent, our own upcoming Agent harness, and other external frameworks.
 
 MinimalAgent adapter support is intentionally not included in v0.4.
 
@@ -293,6 +312,7 @@ tests/                  # Standard-library unittest suite
 - [ ] Add executable benchmark runners for external checkouts.
 - [ ] Add richer rewrite policies and adapter examples.
 - [ ] Add adapters for more agent harnesses after the GenericAgent boundary stabilizes.
+- [ ] Release our own Agent harness for Bayesian-Agent; current experiments use GenericAgent as the backend harness.
 
 
 ## 📈 Star History
@@ -305,7 +325,7 @@ If you use Bayesian-Agent in your research or projects, please cite it as:
 
 ```bibtex
 @software{bayesian_agent_2026,
-  title = {Bayesian-Agent: Bayesian Self-Evolving Agent Framework},
+  title = {Bayesian-Agent: A Bayesian Self-Evolving Agent Framework with Cross-Harness Adaptation},
   author = {{Xiaojun Wu}},
   year = {2026},
   url = {https://github.com/DataArcTech/Bayesian-Agent}

README_ZH.md

@@ -1,4 +1,4 @@
-# Bayesian-Agent
+# Bayesian-Agent: A Bayesian Self-Evolving Agent Framework with Cross-Harness Adaptation
 
 <div align="center">
   <img src="assets/banner.png" width="920" alt="Bayesian-Agent banner"/>
@@ -10,13 +10,19 @@
   <a href="https://github.com/DataArcTech/Bayesian-Agent">GitHub</a>
 </p>
 
-Bayesian-Agent 是一个独立的 Bayesian Self-Evolving Agent 框架，用于把 Agent 的失败轨迹转化为可复用、可验证、带 posterior 权重的 Skills 和 SOPs。
+Bayesian-Agent 是一个面向跨 Agent framework / execution harness 的 Bayesian self-evolving layer，用于把 Agent 的失败轨迹转化为可复用、可验证、带 posterior 权重的 Skills 和 SOPs。
+
+它不是又一个封闭的 monolithic agent framework，而是突出三件事：
+
+- **从零自进化**：没有历史 traces 也能从完整 benchmark 或生产任务中在线沉淀 Skills。
+- **增量修复**：接到已有 Agent 后面，读取失败轨迹，只重跑需要修复的任务。
+- **跨 harness 适配**：当前适配 GenericAgent，后续可通过统一 trajectory schema 和 adapter boundary 接入其他 agent frameworks。
 
 > v0.4 是第一个独立开源版本。它包含 Bayesian Skill Evolution 核心包、Schema、CLI 工具、实验 artifacts，以及干净的 GenericAgent 集成边界。GenericAgent 本身不会被复制、vendoring 或 fork 到本仓库中。
 
 ## 📅 News
 
-- **2026-05-09:** 发布 Bayesian-Agent v0.4 独立 package，包含 Bayesian Skill Evolution 核心 primitives、schemas、CLI utilities 和实验 artifacts。
+- **2026-05-09:** 发布 Bayesian-Agent v0.4 独立 package，包含跨 harness Bayesian Skill Evolution 核心 primitives、schemas、CLI utilities 和实验 artifacts。
 - **2026-05-09:** 增加可选 GenericAgent adapter boundary，不复制、不 vendoring GenericAgent。
 - **2026-05-09:** 发布中英文项目文档和 Bayesian-Agent 方法框架图。
 
@@ -38,12 +44,12 @@ Prompt 可以改善单轮回答。Context 可以改善单次决策。Harness Eng
 - 哪些失败模式要规避
 - 什么时候停止、重试或重写流程
 
-Bayesian-Agent 想回答的问题是：既然 Skill 本质上是“如何完成任务”的假设，为什么它的进化要靠经验堆叠，而不是靠证据更新？
+Bayesian-Agent 想回答的问题是：既然 Skill 本质上是“如何完成任务”的假设，为什么它的进化要靠经验堆叠，而不是靠证据更新？我们的答案是一个框架无关的进化层：既能从零 bootstrap Skills，也能增量修复已有 Agent，还能在不同 harness 之间迁移，只要它们能产出 verified trajectories。
 
 <div align="center">
   <img src="assets/bayesian_agent_overview.png" width="900" alt="Bayesian-Agent overview"/>
   <br/>
-  <em>Bayesian-Agent 将经过验证的 Agent 轨迹转化为 posterior 加权的 Skills 和 SOPs。</em>
+  <em>Bayesian-Agent 将任意兼容 harness 的 verified trajectories 转化为 posterior 加权的 Skills 和 SOPs。</em>
 </div>
 
 ## 🧠 核心思想
@@ -83,8 +89,9 @@ P(success | theta, C, skill)
 - **Bayesian Skill Registry**：维护 Beta posterior、失败模式、token 成本、延迟、轮次和 context 分布。
 - **面向失败模式的修复**：识别反复出现的错误，生成聚焦的 repair plan。
 - **Token-aware context 构建**：只注入真正有 posterior 证据价值的 Skill。
-- **全量与增量两种模式**：既可以从零进化，也可以作为既有 Agent 的修复层。
-- **框架无关的边界**：通过 adapters 集成 GenericAgent 或其他 harness，而不是复制它们的代码。
+- **从零全量自进化**：完整运行任务，在线收集 evidence，并在无历史 traces 的情况下进化 Skills。
+- **已有 Agent 的增量修复层**：读取 baseline agent 的失败轨迹，只重跑失败任务。
+- **跨 harness 适配**：当前集成 GenericAgent，后续通过 adapters 接入其他 agent frameworks，而不是复制它们的代码。
 - **标准库优先**：v0.4 核心运行时不依赖 Python 标准库之外的包。
 
 ## 🧬 自我进化机制
@@ -197,7 +204,7 @@ skill_context = SkillContextBuilder(registry).render(task_context="sop_bench")
 print(skill_context)
 ```
 
-## 🔁 两种运行模式
+## 🔁 三种运行形态
 
 ### 🌱 全量 Self-Evolving Mode
 
@@ -215,6 +222,16 @@ Base Agent -> Failure Traces -> Bayesian Skill Evolution -> Rerun Failures -> Hi
 
 这是更推荐的生产路径，因为它不需要重新训练模型，也不需要替换原有 harness。
 
+### 🔌 Cross-Harness Adaptation Mode
+
+Bayesian-Agent 不绑定某一个 agent runtime。任何 agent framework 只要能产出统一 trajectory schema，并通过 adapter 接收 posterior 加权的 Skill context，都可以成为 Bayesian-Agent 的后端。
+
+```text
+Any Agent Harness -> Trajectory Schema -> Bayesian Skill Registry -> Adapter -> Next Harness Run
+```
+
+这让 Bayesian-Agent 更像一个可移植的 Skill/SOP 进化层，而不是又一个封闭 agent framework。
+
 ## 📊 实验结果
 
 v0.4 原型基于 GenericAgent 与 `deepseek-v4-flash`，在 SOP-Bench 和 Lifelong AgentBench 上完成验证。
@@ -247,13 +264,13 @@ v0.4 原型基于 GenericAgent 与 `deepseek-v4-flash`，在 SOP-Bench 和 Lifel
 | SOP-Bench | GA+BayesianIncremental | deepseek-v4-flash | 100% | 216k | 10k | 226k | 17.73 |
 | Lifelong AgentBench | GA+BayesianIncremental | deepseek-v4-flash | 100% | 71k | 7k | 78k | 25.57 |
 
-这说明 Bayesian-Agent 可以作为即插即用的 repair layer：接在一个未达到 100% 准确率的 Agent 后面，用较小的增量推理成本把失败任务补齐。
+这说明 Bayesian-Agent 可以作为即插即用的 repair layer：接在一个未达到 100% 准确率的 Agent 后面，用较小的增量推理成本把失败任务补齐。这也是它区别于普通 benchmark agent 的关键：它可以站在 harness 旁边，学习它的失败，并在不替换它的情况下提升它。
 
 实验 artifacts 位于 [`artifacts/`](artifacts/)，方法说明位于 [`docs/method.md`](docs/method.md)。
 
-## 🔌 与 GenericAgent 的关系
+## 🔌 GenericAgent 与跨 Harness 适配
 
-第一个原型是在 GenericAgent 内部验证的，但 Bayesian-Agent 不是 GenericAgent fork。
+第一个原型是在 GenericAgent 内部验证的，但 Bayesian-Agent 不是 GenericAgent fork，也不只是 GenericAgent 的附属模块。
 
 开源结构是：
 
@@ -263,7 +280,9 @@ v0.4 原型基于 GenericAgent 与 `deepseek-v4-flash`，在 SOP-Bench 和 Lifel
 - `schemas/`：可移植的 trajectory 与 Skill belief schema
 - `artifacts/`：可复现实验结果文件
 
-GenericAgent 仍然只是可选后端。其他 Agent harness 只要能产出统一 trajectory schema，也可以接入 Bayesian-Agent。
+GenericAgent 是当前实验后端。其他 Agent harness 只要能产出统一 trajectory schema，并实现 adapter boundary，也可以接入 Bayesian-Agent。
+
+长期方向是让 Bayesian-Agent 成为多个 agent runtime 共享的 Bayesian Skill/SOP evolution layer：包括 GenericAgent、我们后续会上传的自研 Agent harness，以及其他外部框架。
 
 MinimalAgent adapter 在 v0.4 中按计划暂不提供。
 
@@ -293,10 +312,11 @@ tests/                  # Standard-library unittest suite
 - [ ] 增加可在外部 checkout 中执行的 benchmark runners。
 - [ ] 增加更丰富的 rewrite policies 和 adapter examples。
 - [ ] GenericAgent 边界稳定后再扩展更多 agent harness adapters。
+- [ ] 上传我们自己的 Agent harness；当前实验阶段使用 GenericAgent 作为 backend harness。
 
 ## 🚦 当前状态
 
-Bayesian-Agent v0.4 是早期独立版本。当前 package 可用于 trace ingestion、Bayesian Skill belief update、context rendering、repair planning 和 result summarization。完整 benchmark execution 仍依赖 GenericAgent 等外部 Agent harness。
+Bayesian-Agent v0.4 是早期独立版本。当前 package 可用于 trace ingestion、Bayesian Skill belief update、context rendering、repair planning 和 result summarization，并已经验证从零全量进化、增量修复、跨 harness 适配三条路径。完整 benchmark execution 当前仍使用 GenericAgent 等外部 Agent harness；后续会上传我们自己的 Agent harness。
 
 ## 📈 Star History
 
@@ -308,7 +328,7 @@ Bayesian-Agent v0.4 是早期独立版本。当前 package 可用于 trace inges
 
 ```bibtex
 @software{bayesian_agent_2026,
-  title = {Bayesian-Agent: Bayesian Self-Evolving Agent Framework},
+  title = {Bayesian-Agent: A Bayesian Self-Evolving Agent Framework with Cross-Harness Adaptation},
   author = {{Xiaojun Wu}},
   year = {2026},
   url = {https://github.com/DataArcTech/Bayesian-Agent}

docs/adapters.md

@@ -1,6 +1,20 @@
 # Adapters
 
-Bayesian-Agent is designed to integrate with external agent harnesses without copying their code.
+Bayesian-Agent is designed to integrate with external agent harnesses without copying their code. This is one of the main reasons the project is not just another agent framework: the Bayesian layer can improve whichever harness emits verified trajectories.
+
+## Adaptation Advantage
+
+Bayesian-Agent separates Skill evolution from task execution:
+
+```text
+Harness executes -> Bayesian-Agent learns -> Adapter injects posterior Skill context -> Harness reruns
+```
+
+That separation enables three deployment styles:
+
+- run a full benchmark from scratch with Bayesian Skill evolution enabled
+- repair only the failed tasks from an existing agent run
+- reuse the same Skill belief registry across compatible harnesses
 
 ## Adapter Contract
 
@@ -50,6 +64,10 @@ External systems should emit:
 
 Bayesian-Agent can then update beliefs and render the next Skill context.
 
+## Planned Bayesian-Agent Harness
+
+Current experiments use GenericAgent as the backend harness. A dedicated Bayesian-Agent harness is planned so users can run the full loop without depending on GenericAgent, while still keeping GenericAgent and other frameworks as optional backends.
+
 ## MinimalAgent Status
 
 MinimalAgent adapter support is intentionally not included in v0.4.
@@ -58,4 +76,5 @@ The recommended path is:
 
 1. stabilize the GenericAgent boundary
 2. keep the core trace schema portable
-3. add more adapters only after the adapter contract has enough real usage
+3. upload the dedicated Bayesian-Agent harness
+4. add more adapters only after the adapter contract has enough real usage