What does this skill do?

The IA Skill Agent Architecture provides a comprehensive framework for designing and optimizing the scope of action for AI agents: it defines tools without overlap, standardizes the format of observations with required retrieval fields, and measures the actual completeness rate. It audits existing agents, classifies tool granularity, and selects the optimal architectural pattern (ReAct, function-calling, or hybrid) based on the use case.

Audit of Scope of Action
It detects semantic overlaps and overly broad "catch-all" tools, and refactors the tool inventory to eliminate ambiguities.
Granularity Classification
Assign each tool to micro (high risk), medium (loops), or macro (round-trip cost), and optimize the context budget.
Recovery Contracts
For each error path, define the root cause, safe retry instruction, and explicit stop condition, thereby reducing infinite loops.
Performance Metrics
It measures the completion rate, retries per task, pass@1, pass@3, and cost per successful task to iterate using real data.

Usage examples

🔍 Comprehensive audit
Check the scope of my deployment agent: I have 7 tools, and the completion rate is 62%. Detect overlaps and suggest refactoring.
📋 Definition of a tool
Create a micro-granularity tool to deploy services into production with a full recovery contract.
📊 Pattern Comparison
Compare ReAct, function-calling, and hybrid approaches for an agent that manages cloud infrastructure. Which one maximizes the success rate?
🛡️ Error Contracts
My agent gets stuck in loops when permissions fail. Define the recovery contract with root cause, safe retry, and stop condition.

Features

Tabular Inventory of Tools Classify each tool by granularity, detect overlaps, and generate specific refactoring actions.
Standardized Observation Protocol Every tool response includes status, summary, next_actions, and artifacts, as well as error recovery fields.
A Comparison of Architectural Patterns Choose between ReAct (exploratory), function-calling (deterministic), or hybrid, depending on the use case and budget.
Performance Metrics It measures the completion rate, retries per task, pass@1/pass@3, and cost per successful task to iterate through the data.
Context Optimization Keep system prompts to a minimum, move guides to on-demand skills, and group them by phase boundaries, not arbitrary thresholds.

Frequently asked questions

The set of tools available to the agent. This skill helps you audit them, eliminate overlaps, and classify each tool by granularity (micro/medium/macro) to maximize the success rate.
ReAct for exploratory tasks with an uncertain path (e.g., research, debugging). Function-calling for structured deterministic workflows (e.g., data pipelines). Hybrid for planning with ReAct + execution with typed tools.
For each error path, document the likely root cause, the safe retry instruction, and the explicit stop condition. This prevents infinite loops and improves the completion rate.
Using four metrics: completion rate (% of tasks completed), retries per task, pass@1 and pass@3 (success on the first/third attempt), and cost per successful task. If the completion rate does not exceed 70%, the problem usually lies in the scope of action.
AI Agent Architecture — Design and Optimization of Agents with Claude AI

¿Prefieres escuchar el contenido? Genera la narración de audio con un clic.