agent-data-protocol

name	agent-data-protocol
title	Agent Data Protocol: Unifying Datasets for Diverse, Effective Fine-tuning of LLM Agents
version	0.0.2
engine	skillxiv-v0.0.2-claude-opus-4.6
license	MIT
url	https://arxiv.org/abs/2510.24702
keywords	["Data Format","Fine-tuning","Agent Training","Standardization","Interoperability"]
description	Standardizes agent training data representation across diverse sources (API use, web browsing, coding, software engineering). Single lightweight protocol unifies 13 datasets enabling 20% performance gains without domain-specific tuning. Enables reproducible agent training and scalable data combination.

Agent Data Protocol: Unified Training Data Format

Agent training suffers from dataset fragmentation: different sources use incompatible formats (ReAct, OpenAI, custom). ADP provides lightweight, expressive standard format that unifies diverse datasets into single training format.

A single model trained on unified data outperforms task-specific variants, enabling scalable agent training.

Core Concept

Key innovation: single interlingua that captures diverse agent interactions:

• API calls, tool use, planning, execution
• Web browsing, DOM interaction, form filling
• Code generation, execution, debugging
• Generic enough for new domains without redesign

Format is simple to parse and extensible for new agent capabilities.

Architecture Overview

• Minimal schema capturing agent interactions
• Tool/API invocation representation
• Observation and action specification
• Task metadata and success signals

Implementation Steps

Define the core ADP schema. Keep it minimal to reduce parsing overhead:

from dataclasses import dataclass
from typing import Dict, List, Any

@dataclass
class ADPAction:
    """Unified agent action representation."""
    type: str  # "tool", "code", "text", "navigate"
    tool_name: str = None  # e.g., "search", "click", "execute"
    parameters: Dict[str, Any] = None  # tool parameters
    code: str = None  # if type=="code"
    text: str = None  # if type=="text"

@dataclass
 :
    
    status:   
    output:   
    metadata: [, ] =   


 :
    
    action: ADPAction
    observation: ADPObservation
    reward:  =   


 :
    
    task_instruction: 
    steps: [ADPStep]
    final_success: 
    domain: