Project archive / 01
Enterprise Long-Horizon Agent System
Z-Claw
A long-horizon agent system for enterprise and office workflows, supporting complex tasks, business system integrations, browser operations, scheduled jobs, and multi-skill execution.
- Archive
- 01
- Category
- Enterprise Long-Horizon Agent System
- Result
- 97% core-task regression success rate
System / highlights
What I delivered.
- 01
Designed and implemented a custom Agent Runtime with Fast and Expert routing modes for tasks of different complexity.
- 02
Built a skill-routing pipeline combining BM25 matching, vector retrieval, reranking, and RAG to map user intent to reusable Skill APIs.
- 03
Defined Skill schemas, invocation protocols, parameter constraints, and execution feedback through multiple rounds of real-user validation.
- 04
Created an Agent Eval framework for regression testing and version iteration, achieving a 97% success rate on core tasks.
- 05
Designed a multi-layer Context Manager and memory recall system that scored 78 on MemBench.
- 06
Introduced isolated sandboxes for browser automation and lightweight tool execution, plus administration and scheduled-task services.
Archive / full note
Project notes.
A production long-horizon agent system for complex enterprise and office workflows.
Overview
Z-Claw is a long-horizon agent system designed for real enterprise work rather than isolated demonstrations. It coordinates reusable skills, internal systems, browser operations, scheduled jobs, and multi-step execution while keeping context and failures observable.
The runtime routes straightforward work through a fast path and sends ambiguous or high-complexity tasks through a deeper expert path. Both modes share the same governed skill contracts and execution feedback, allowing improvements to be measured instead of relying on anecdotal prompt changes.
Retrieval and skill routing
User intent is mapped to reusable skills through a retrieval pipeline that combines BM25, vector search, reranking, and RAG. Each skill exposes a defined schema, parameter constraints, and execution contract so the agent can reason about capabilities without relying on hidden prompt conventions.
Reliability
Regression evaluation is part of the product loop. Core scenarios are replayed across runtime and prompt versions, while LangSmith traces and execution feedback expose where planning, retrieval, parameter generation, or tool execution failed. Sensitive implementation and business details remain intentionally anonymized.
Index / methods
Technology & methods.
- 01Agent Runtime
- 02BM25
- 03Vector Search
- 04Reranking
- 05RAG
- 06Context Manager
- 07Cube Sandbox
- 08LangSmith