Can the agent correct its own mistakes?

The skill includes a 'Self-Correction Loop' pattern where the LLM validates its own output against specific criteria and retries with refined instructions if the initial result is invalid.

How does this skill handle long conversations?

It uses a sliding window memory management system that automatically compresses and summarizes older messages to prevent context overflow while retaining essential history.

What is the ReAct pattern implemented in this skill?

The ReAct pattern combines reasoning ('Thought') and acting ('Action') in an iterative loop, allowing the LLM to process observations from tools before deciding on the next logical step.

Does it prevent agents from getting stuck in infinite loops?

Yes, the patterns include configurable 'max_steps' limits (recommended 5-15) to ensure that the agent terminates or requests human intervention if a solution isn't found.

Agentic Workflows & Loops

Name: Agentic Workflows & Loops
Author: yonatangross

byyonatangross

•

Data Science & ML

Implements autonomous reasoning patterns like ReAct and Plan-and-Execute to enable LLMs to solve complex, multi-step tasks.

This skill provides production-ready templates and implementation patterns for building sophisticated AI agents capable of autonomous reasoning. It includes industry-standard architectures such as ReAct (Reasoning + Acting), Plan-and-Execute cycles, and self-correcting loops. By incorporating advanced memory management with sliding window context handling, it allows developers to create agents that manage long-running tasks, recover from execution errors, and maintain consistency across complex workflows without exceeding context limits.

Key Features

01Sliding window memory management with context compression and summarization

02Plan-and-Execute architecture for structured, multi-step task completion

03ReAct (Reasoning + Acting) pattern implementation for iterative problem solving

04Self-correction loops for output validation and automatic error recovery

05Configurable execution parameters for step limits, temperature, and validation frequency

0669 GitHub stars

Use Cases

01Building autonomous research agents that use tools to find, verify, and synthesize information

02Creating coding assistants that plan, implement, and self-correct based on error logs

03Developing complex multi-step automation where the AI must adapt based on real-time tool observations

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add yonatangross/orchestkit agent-loops

For use in Claude.ai and ChatGPT

Download Skill