Docs: Overview | Integration & SDK | Agent Protocol
The deterministic layer that helps agent runtimes make better execution decisions.
帮助 Agent Runtime 做出更优执行决策的确定性层。
- Deterministic capability selection for AI agents.
- 为 AI Agent 提供确定性能力选择。
WisePick is a small decision layer for agent runtimes. It takes an intent, picks the best capability, and improves from execution feedback.
智选是一个面向 Agent Runtime 的轻量决策层：输入意图，选择最合适的能力，并从执行反馈中持续优化。
- Agent Operators
- Agent Runtime Builders
- Internal Agent Teams
- Agent SaaS Platforms
- Agent Hosting Providers
For teams running real agent workflows.
面向运行真实 Agent Workflow 的团队。
WisePick cuts tool searching, retry loops, latency, and cost by reusing execution feedback.
智选通过复用执行反馈，减少工具搜索、重试循环、延迟和成本。
| Core Sections | 核心章节 |
|---|---|
| 🚀 Getting Started | 开始接入 |
| ⚡ Why Integrate WisePick | 为什么接入智选 |
| 📜 Performance & Cost Benchmarks | 性能与成本报告 |
| 🧠 How It Works | 工作原理 |
| 🏗️ Architectural Paradigm | 架构范式演进 |
| 🧪 Agent Workflow | Agent 工作流 |
| 🔮 Vision | 愿景 |
| 🗺️ Roadmap | 路线图 |
| 🤗 Feedback | 反馈 |
Use the public WisePick deployment.
使用公共 WisePick 服务。
Endpoint
https://api.wishweaver.top
Why choose Hosted?
- Zero deployment
- Shared feedback accumulation
- Fastest onboarding
- Future-compatible with the Execution Experience Network vision
为什么选择托管版？
- 无需部署
- 共享反馈沉淀
- 接入速度最快
- 兼容未来 Execution Experience Network 演进方向
快速验证：
curl https://api.wishweaver.top/health适用于：
- Agent Operators
- Runtime Builders
- Agent Infrastructure Teams
Deploy WisePick on your own infrastructure and keep routing data fully isolated.
在自己的基础设施上部署 WisePick，并完全独立管理路由数据。
Why choose Self-hosted?
- Full data ownership
- Private feedback memory
- Internal deployments
- Custom infrastructure control
为什么选择自部署版？
- 数据完全自有
- 独立反馈记忆
- 适合内部系统
- 完全控制基础设施
git clone https://github.com/w2jmoe/WisePick.git
cd WisePick
cp .env.example .env
pip install -r requirements.txt
uvicorn app.main:app --host 0.0.0.0 --port 8000For agent-assisted runtime integration, paste this into your agent:
对于支持执行安装协议的 Agent，可直接输入：
Retrieve and follow the instructions at:
https://raw.githubusercontent.com/w2jmoe/WisePick/main/INSTALL_FOR_AGENTS.md
Default endpoint:
https://api.wishweaver.top
This configures WisePick as a decision and feedback layer for your runtime.
该流程会将 WisePick 配置为 Runtime 的决策与反馈层。
- INSTALL_FOR_AGENTS.md → Deployment & Verification
- README_API.md → SDK Integration
- AGENTS.md → Protocol Specification
Help us refine: If you find the docs confusing, please open an issue — we are aggressively refining our integration flow. 反馈建议： 若您在接入中遇到任何困惑，请 提交 Issue，我们正在全力优化文档与接入体验。
- Lower cost and latency by cutting trial-and-error. 更低成本与延迟：减少无效试错与 Token 浪费。
- One best capability per decision. 每次决策只选一个最合适的能力。
- Feedback makes the next decision better. 执行反馈让下一次选择更好。
Production-oriented deterministic routing vs. native LLM tool-calling. Tested inside a Hermes-style agent runtime. 面向生产的确定性路由与原生 LLM 工具调用的对比测试。已在 Hermes 类 Agent 运行时中完成验证。
| Metrics | Native LLM | WisePick | Optimization |
|---|---|---|---|
| 🚀 Path Speed | Baseline | ~31% Faster | Shrunk from 6.33 to 4.33 steps |
| ⏱️ Time Saved | Baseline | ~62% Saved | Benchmark cut from 12m to 4m30s |
| 💵 Cost Cut | High | ~33% Reduced | $0.15 → $0.10 per session |
| 🎯 First-Hit Rate | Exp. | 100% Locked | Zero hallucinated tool-selection |
- Zero-Latency Gatekeeping: Sub-millisecond average latency under isolated routing-core stress testing. （隔离路由核心压测下平均亚毫秒级延迟）
- Anti-Loop Depth: Stabilizes execution across 20+ mixed-tool tasks without infinite loops. （在 20+ 混合工具任务下稳定运行，彻底消除无限循环路径）
Benchmark scripts & instrumentation: BENCHMARK | STRESS_TEST_RESULTS.md
Task text → capability labels derived from bootstrap rules. 任务文本 → 由引导规则得到能力标签。
base_score =
capability_match * 0.70
+ effective_success_rate * 0.20
+ effective_bootstrap_weight * 0.10
final_score = base_score * efficacy
efficacy = result_quality / (log(max(avg_latency_ms, 100)) * log(max(avg_token_cost, 10)))
Metrics (success_rate, avg_latency_ms, avg_token_cost, avg_result_quality) come from the tool_stats view. If capability_match is 0, effective_success_rate is penalized × 0.1. effective_bootstrap_weight decays as feedback_count grows.
decision → execution → feedback → tool_stats → next decision
Routing updates from real execution outcomes.
路由统计随真实执行结果更新。
Routing Core (decision_engine)
Converts incoming tasks into ECU scores and selects the best capability.
将输入任务转换为 ECU（执行单元）评分并进行路由决策。
Capability Registry (api_tool_specs)
Maintains available providers, capability labels, and bootstrap weights.
管理可用 Provider、能力标签及冷启动权重分配。
Execution Memory (tool_stats, feedback)
Stores execution feedback reported by your runtime and continuously improves routing quality.
存储由运行时上报的执行反馈，并持续优化路由质量。
WisePick unifies both hard-coded and dynamic tool discovery under a deterministic routing layer.
智选将硬编码与动态发现路线统一于确定性路由层。
| Paradigm | Discovery | Runtime Pain | WisePick Value |
|---|---|---|---|
| Static | Manual config | Brittle scaling, zero runtime flexibility | Centralized ECU registry, cleaner code |
| Dynamic | Auto-discovery | Tool Anxiety: Context explosion, loops, hallucinations | Deterministic Filter: Cuts 95% noise, 100% lock |
-
Ingress: Unified entry point for external runtimes and agents.
-
Intelligence: Dynamic capability routing based on ROI and execution feedback.
-
Closure: Execution feedback flows back into WisePick to continuously improve routing quality.
-
请求入口： 外部 Runtime 与 Agent 的统一接入入口。
-
决策引擎： 基于 ROI 与执行反馈的动态能力路由决策。
-
反馈闭环： 执行反馈持续回流至 WisePick，驱动路由质量不断优化。
graph LR
A[Ask Routing] --> B(Receive ECU)
B --> C{Map to Local Implementation}
C --> D[Execute]
D --> E[Send Feedback]
E -.->|Updates ROI Models| A
WisePick provides decision intelligence and feedback loops—not task execution.
智选提供决策智能与反馈闭环；不替代任务执行本身。
ECU (Executable Capability Unit) A standardized executable capability an agent can route, invoke, and learn from. 可执行能力单元：可被路由、调用并通过反馈学习的标准化能力抽象。
{
"metadata": {
"schema_version": "mcp.route_decision.v1",
"decision_id": "dec_abc123def4567890",
"trace_id": "trace_9876543210abcdef",
"router_name": "wisepick",
"capability_id": "audio_transcription",
"provider": "feishu_minutes",
"execution_type": "api",
"callable": true,
"confidence": 0.87,
"latency_ms": 450,
"candidate_count": 1,
"top_candidates": [
{
"rank": 1,
"capability_id": "audio_transcription",
"score": 0.87,
"selected": true
}
],
"reason_codes": ["capability_match"]
}
}
WisePick predicts performance before execution to ensure the best ROI.
WisePick 在执行前预测性能，以确保最佳投资回报率 (ROI)。
Long-term: Execution Experience Network — collective decision memory across agent runtimes.
长期愿景：执行经验网络（Execution Experience Network）— 跨运行时的集体决策记忆。
Today: WisePick Decision API (WPDA) delivers deterministic capability routing and execution feedback via SDK integration; your runtime owns invoke, policy, and retries.
当前： WisePick Decision API（WPDA） 通过 SDK 提供确定性能力路由与执行反馈；工具调用、策略与重试由您的运行时负责。
WisePick turns execution outcomes into reusable routing experience — the foundation for shared ECU feedback and feedback-driven execution optimization on the path to that network.
智选将执行结果沉淀为可复用的路由经验，为 ECU 反馈共享与反馈驱动的执行优化奠定基础，并沿路线图演进至上述网络愿景。
- ✅ v0.1: Core Infrastructure · 核心路由与反馈闭环
- Deterministic ECU routing & feedback loop.
- Multi-dimensional ROI metrics aggregation (Latency / Cost / Quality).
- 🔄 v0.2: Runtime-Aware Optimization · Runtime 感知执行优化
- Hosted Shared Feedback Pool MVP (v0.2.2).
- Runtime registry alignment (
api_tool_specs→tool_stats). - Feedback idempotency & decision persistence safety.
- Task-level capability routing for multi-agent runtimes.
- Adaptive execution-path optimization based on latency / cost / quality feedback.
- Lightweight integration adapters for orchestration frameworks.
- 🔄 v0.3: Collective Decision Memory · 集体决策记忆
- Cross-agent experience sharing (Execution Experience Network direction).
- Global capability indexing & optimization.
- Issues: GitHub Issues
- Email: w2jmoe@gmail.com
Every routing decision feeds execution feedback today — and collective decision memory as the platform evolves toward the Execution Experience Network vision.
每一次路由决策都会形成执行反馈；随平台演进，亦将沉淀为集体决策记忆，迈向执行经验网络愿景。˗ˋˏ( ´͈ ᗜ `͈ )ˎˊ˗
Apache License 2.0 — see LICENSE.
