holny/agent-harness-develop-book
Developer toolsThis page was translated automatically. Technical details may differ from the original. View the original
Complete Agent Harness Architecture and Core Module Development Guide - Reference to production-grade agent source code: Claude Code, OpenCode, OpenClaw, and Hermess
Overview
Agent Harness Dev Book (Agent Harness Development Guide) — References to industry-level production-grade agent source code: Claude Code, OpenCode, OpenClaw, and Hermes. If you find it useful, please give it a little star and share it with everyone. Thank you This definition is from a technical implementation perspective, pointing out that to build an Agent, it needs to be based on large models, needs planning capabilities, can think about what to do next, needs memory—including long-term and short-term memory—and needs to be able to use tools. The combination of these abilities is called an Agent. The value of this formula lies in **. 1. In October 2022, the theoretical cornerstone of Agent ReAct was proposed: a ReAct paper jointly published by a team from Princeton University and Google Research titled "ReAct: Synergizing Reasoning and Acting in Language Models," which centers on enabling large models to interact while reasoning through a closed-loop process of "Thought→ Action, → Observation." 2. In June 2023, proposed the Agent definition (Four-Element Formula): This formula was systematically explained by OpenAI's Head of Applications Research in the well-known blog post "LLM Powered Autonomous Agents" published in June 2023. In the field of AI engineering, it is borrowed and defined as:** Its purpose is: to enable large models with non-deterministic and probabilistic characteristics to behave as deterministic, reliable, and controllable "workers" in production environments. The original meaning of 'harness' in English is "horse gear"—a leather and metal device worn on the horse's body to control direction, bear heavy loads, and connect the carriage. Harness concept proposal timeline: 1. Concept launch in November 2025: In its engineering blog "Effective harnesses for long-running agents," the Claude Agent SDK was first described as "a powerful, general-purpose Agent Harness." 3.
Overview (machine translated)
Agent Harness Dev Book
—— Based on production-grade agent source code: Claude Code, OpenCode, OpenClaw, Hermes, DeepSeek Harness, Codex, and Pi
If you find this useful, please give it a star and share it with others — thank you!
English | Chinese
📕 PDF version (viewable online on GitHub): Agent Harness Explained (English) | Agent Harness Detailed Explanation (Chinese)
2026-09 changelog (based on 20 engineering articles + deep source-code reading of 7 projects):
| # | Added / updated | Highlights |
|---|---|---|
| 1 | DSH (DeepSeek Harness) source-code supplement (folded into §1/§2/§3/§4/§5/§6) | Cordis plugin kernel, event-sourcing log, six-state TurnEndReason, two-level compaction + Spill, three-stage progressive Skills, SubAgent 6 providers, Code Mode run_code, guard pipeline (repeat-tool-reminder + timeout-policy at source level) |
| 2 | Codex (Rust) source-code supplement (folded into §1/§3/§4/§5/§6) | Thread/Turn/Item three layers, independent Approval + Sandbox layers, two-stage memory pipeline (stage-1 extraction → stage-2 consolidation + git baseline + subagents), 90%-threshold dual-scope compaction |
This page was translated automatically. Technical details may differ from the original.
Recommended Tools
Try a different keyword or remove a filter.
Install
npx skillfish add holny/agent-harness-develop-book