When should I use a sub-agent versus a context fork?

Use 'context: fork' for most isolation needs as it is significantly more cost-effective. Sub-agents should only be utilized when the specific benefits of parallelization clearly exceed the 20K token startup cost and associated quota overhead.

How does this skill handle context window pressure?

It provides patterns for context offloading, such as writing tool results to files and using summarization, alongside context caching strategies to maintain a stable prefix and reduce latency.

What is the Progressive Disclosure pattern?

Progressive Disclosure is a strategy where the agent is provided a high-level index of capabilities and only retrieves full tool specifications or documentation just-in-time when a specific task requires that knowledge, saving context space.

What are the benefits of the Multi-Layer Action Space?

By organizing actions into hierarchies—from tool calling to shell utilities and generated scripts—you reduce the overhead of tool definitions while maintaining granular control over the agent's environment.

Architecting Agents

Name: Architecting Agents
Author: Git-Fg

byGit-Fg

•

開発者ツール

Implements industry-proven design patterns for building high-performance AI agents and optimizing context management systems.

概要

Architecting Agents provides a standardized framework for designing robust AI agent behaviors based on production systems like Claude Code and Cursor. It offers comprehensive guidance on optimizing context windows through offloading and caching, implementing hierarchical tool layers via progressive disclosure, and making cost-effective architectural decisions regarding sub-agent isolation versus forking. By following these established protocols, developers can build more reliable, scalable, and cost-efficient agentic workflows while avoiding common pitfalls like context saturation or excessive token consumption during complex development tasks.

主な機能

Evolutionary memory systems for distilling session trajectories into reusable skills
Cost-optimized context isolation patterns using forks instead of high-cost sub-agents
Persistent state management patterns via filesystem and shell utility integration
1 GitHub stars
Hierarchical tool design to reduce model overhead and improve execution accuracy
Just-in-time context loading through progressive disclosure of tool specifications

ユースケース

Implementing 'Ralph Loops' for iterative task fulfillment and automated progress verification
Designing complex multi-agent workflows that require parallel execution without excessive token costs
Managing massive context windows for long-running development tasks in large-scale codebases

概要

主な機能

Evolutionary memory systems for distilling session trajectories into reusable skills
Cost-optimized context isolation patterns using forks instead of high-cost sub-agents
Persistent state management patterns via filesystem and shell utility integration
1 GitHub stars
Hierarchical tool design to reduce model overhead and improve execution accuracy
Just-in-time context loading through progressive disclosure of tool specifications

ユースケース

Implementing 'Ralph Loops' for iterative task fulfillment and automated progress verification
Designing complex multi-agent workflows that require parallel execution without excessive token costs
Managing massive context windows for long-running development tasks in large-scale codebases