Self-Evolving Agents

🔄

Self-Evolving Loop

Automated feedback loop that captures issues, learns from feedback, and promotes improvements back into production workflows.

🤖

GPT-4.1 Agentic Prompting

Learn the exact prompts used to achieve top SWE-bench scores including persistence, reflection, and planning patterns.

🛡️

Production Guardrails

Input/output safety guardrails with async parallel execution for minimal latency impact in production systems.

⚡

Performance Optimization

Prompt caching (80% faster), batch processing (50% cheaper), and fine-tuning techniques for production scale.

🧪

Four Complementary Graders

Balance deterministic checks (chemical names, length) with semantic judgment (cosine similarity, LLM-as-judge).

🏗️

Agent Orchestration

Multi-agent systems with handoffs, routing, swarm patterns, and the official AgentKit framework.

What You'll Learn

┌─────────────────────────────────────────────────────────────────────┐
│                    SELF-EVOLVING AGENT LOOP                         │
├─────────────────────────────────────────────────────────────────────┤
│                                                                      │
│   ┌──────────────┐     ┌──────────────┐     ┌──────────────┐       │
│   │   Baseline   │────►│   Feedback   │────►│    Evals     │       │
│   │    Agent     │     │  (Human/LLM) │     │   & Score    │       │
│   └──────────────┘     └──────────────┘     └──────┬───────┘       │
│          ▲                                          │               │
│          │              ┌──────────────┐            │               │
│          └──────────────│   Updated    │◄───────────┘               │
│                         │    Agent     │    (if score > threshold)  │
│                         └──────────────┘                            │
│                                                                      │
└─────────────────────────────────────────────────────────────────────┘

Documentation Sections

Core Self-Evolving Agents

Section	Description
Quick Start	Get up and running in under 10 minutes
Environment Setup	Install dependencies and configure API keys
Create Evals	Set up the four graders for evaluation
Self-Evolving Loop	Complete automated optimization workflow
GEPA Integration	Advanced genetic-pareto optimization

Section

Description

Quick Start

Get up and running in under 10 minutes

Environment Setup

Install dependencies and configure API keys

Create Evals

Set up the four graders for evaluation

Self-Evolving Loop

Complete automated optimization workflow

GEPA Integration

Advanced genetic-pareto optimization

Advanced Topics (New!)

Section	Description
GPT-4.1 Prompting	Agentic prompts, SWE-bench patterns, CoT
Structured Outputs	JSON schema, Pydantic, strict mode
LLM Guardrails	Input/output safety, async patterns
Function Calling	OpenAPI to functions, tool use
Orchestrating Agents	Multi-agent coordination, handoffs
Parallel Agents	Concurrent execution, DAG workflows
AgentKit Walkthrough	Official OpenAI agent framework

Section

Description

GPT-4.1 Prompting

Agentic prompts, SWE-bench patterns, CoT

Structured Outputs

JSON schema, Pydantic, strict mode

LLM Guardrails

Input/output safety, async patterns

Function Calling

OpenAPI to functions, tool use

Orchestrating Agents

Multi-agent coordination, handoffs

Parallel Agents

Concurrent execution, DAG workflows

AgentKit Walkthrough

Official OpenAI agent framework

Optimization (New!)

Section	Description
Prompt Caching	80% latency reduction, 50% cost savings
Batch Processing	50% cost reduction, async processing
Fine-Tuning Techniques	SFT, DPO, RFT explained
Data Preparation	Validation, token counting, cost estimation
Evaluation Flywheel	Continuous improvement cycles

Section

Description

Prompt Caching

80% latency reduction, 50% cost savings

Batch Processing

50% cost reduction, async processing

Fine-Tuning Techniques

SFT, DPO, RFT explained

Data Preparation

Validation, token counting, cost estimation

Evaluation Flywheel

Continuous improvement cycles

Key Concepts

┌─────────────────────────────────────────────────────────────────────┐
│                    OPTIMIZATION STRATEGIES                           │
├─────────────────────────────────────────────────────────────────────┤
│                                                                      │
│  COST REDUCTION              QUALITY IMPROVEMENT                    │
│  ├─► Batch API (50% off)     ├─► Fine-tuning (SFT, DPO, RFT)       │
│  ├─► Prompt caching          ├─► Prompt optimization                │
│  └─► Smaller models          └─► Evaluation-driven iteration       │
│                                                                      │
│  LATENCY REDUCTION           SAFETY                                 │
│  ├─► Prompt caching (80%)    ├─► Input guardrails                  │
│  ├─► Parallel execution      ├─► Output guardrails                 │
│  └─► Streaming               └─► Async validation                  │
│                                                                      │
└─────────────────────────────────────────────────────────────────────┘

Based On

This documentation consolidates knowledge from the OpenAI Cookbook, including:

Self-Evolving Agents - Bain & OpenAI collaboration

GPT-4.1 Prompting Guide - Agentic best practices

Structured Outputs - Reliable JSON generation

LLM Guardrails - Production safety patterns

Batch API & Caching - Cost optimization

Fine-Tuning Guides - SFT, DPO, RFT techniques

Agent Orchestration - Multi-agent patterns

Original Contributors: Calvin Maguranis, Fanny Perraudeau, Giorgio Saladino, Shikhar Kwatra, Valentina Frenkel

Self-Evolving AgentsA Cookbook for Autonomous Agent Retraining

Self-Evolving Loop

GPT-4.1 Agentic Prompting

Production Guardrails

Performance Optimization

Four Complementary Graders

Agent Orchestration

What You'll Learn ​

Documentation Sections ​

Core Self-Evolving Agents ​

Advanced Topics (New!) ​

Optimization (New!) ​

Key Concepts ​

Based On ​

What You'll Learn

Documentation Sections

Core Self-Evolving Agents

Advanced Topics (New!)

Optimization (New!)

Key Concepts

Based On