erdo 0.1.34

Python SDK for building workflow automation agents with Erdo

Erdo Agent SDK

Build AI agents and workflows with Python. The Erdo Agent SDK provides a declarative way to create agents that can be executed by the Erdo platform.

Installation

pip install erdo

Quick Start

Creating Agents

Create agents using the Agent class and define steps with actions:

from erdo import Agent, state
from erdo.actions import memory, llm
from erdo.conditions import IsSuccess, GreaterThan

# Create an agent
data_analyzer = Agent(
    name="data analyzer",
    description="Analyzes data files and provides insights",
    running_status="Analyzing data...",
    finished_status="Analysis complete",
)

# Step 1: Search for relevant context
search_step = data_analyzer.step(
    memory.search(
        query=state.query,
        organization_scope="specific",
        limit=5,
        max_distance=0.8
    )
)

# Step 2: Analyze the data with AI
analyze_step = data_analyzer.step(
    llm.message(
        model="claude-sonnet-4-20250514",
        system_prompt="You are a data analyst. Analyze the data and provide insights.",
        query=state.query,
        context=search_step.output.memories,
        response_format={
            "Type": "json_schema",
            "Schema": {
                "type": "object",
                "required": ["insights", "confidence", "recommendations"],
                "properties": {
                    "insights": {"type": "string", "description": "Key insights found"},
                    "confidence": {"type": "number", "description": "Confidence 0-1"},
                    "recommendations": {"type": "array", "items": {"type": "string"}},
                },
            },
        },
    ),
    depends_on=search_step,
)

Code Execution with External Files

Use the @agent.exec decorator to execute code with external Python files:

from erdo.types import PythonFile

@data_analyzer.exec(
    code_files=[
        PythonFile(filename="analysis_files/analyze.py"),
        PythonFile(filename="analysis_files/utils.py"),
    ]
)
def execute_analysis():
    """Execute detailed analysis using external code files."""
    from analysis_files.analyze import analyze_data
    from analysis_files.utils import prepare_data

    # Prepare and analyze data
    prepared_data = prepare_data(context.parameters.get("dataset", {}))
    results = analyze_data(context)

    return results

Conditional Step Execution

Handle step results with conditions:

from erdo.conditions import IsSuccess, GreaterThan

# Store high-confidence results
analyze_step.on(
    IsSuccess() & GreaterThan("confidence", "0.8"),
    memory.store(
        memory={
            "content": analyze_step.output.insights,
            "description": "High-confidence data analysis results",
            "type": "analysis",
            "tags": ["analysis", "high-confidence"],
        }
    ),
)

# Execute detailed analysis for high-confidence results
analyze_step.on(
    IsSuccess() & GreaterThan("confidence", "0.8"),
    execute_analysis
)

Complex Execution Modes

Use execution modes for advanced workflows:

from erdo import ExecutionMode, ExecutionModeType
from erdo.actions import bot
from erdo.conditions import And, IsAny
from erdo.template import TemplateString

# Iterate over resources
analyze_files = agent.step(
    action=bot.invoke(
        bot_name="file analyzer",
        parameters={"resource": TemplateString("{{resources}}")},
    ),
    key="analyze_files",
    execution_mode=ExecutionMode(
        mode=ExecutionModeType.ITERATE_OVER,
        data="parameters.resource",
        if_condition=And(
            IsAny(key="dataset.analysis_summary", value=["", None]),
            IsAny(key="dataset.type", value=["FILE"]),
        ),
    )
)

Loading Prompts

Use the Prompt class to load prompts from files:

from erdo import Prompt

# Load prompts from a directory
prompts = Prompt.load_from_directory("prompts")

# Use in your agent steps
step = agent.step(
    llm.message(
        system_prompt=prompts.system_prompt,
        query=state.query,
    )
)

State and Templating

Access dynamic data using the state object and template strings:

from erdo import state
from erdo.template import TemplateString

# Access input parameters
query = state.query
dataset = state.dataset

# Use in template strings
template = TemplateString("Analyzing: {{query}} for dataset {{dataset.id}}")

Core Concepts

Actions

Actions are the building blocks of your agents. Available action modules:

• erdo.actions.memory - Memory storage and search
• erdo.actions.llm - Large language model interactions
• erdo.actions.bot - Bot invocation and orchestration
• erdo.actions.codeexec - Code execution
• erdo.actions.utils - Utility functions
• erdo.actions.resource_definitions - Resource management

Conditions

Conditions control when steps execute:

• IsSuccess(), IsError() - Check step status
• GreaterThan(), LessThan() - Numeric comparisons
• TextEquals(), TextContains() - Text matching
• And(), Or(), Not() - Logical operators

Types

Key types for agent development:

• Agent - Main agent class
• ExecutionMode - Control step execution behavior
• PythonFile - Reference external Python files
• TemplateString - Dynamic string templates
• Prompt - Prompt management

Advanced Features

Multi-Step Dependencies

Create complex workflows with step dependencies:

step1 = agent.step(memory.search(...))
step2 = agent.step(llm.message(...), depends_on=step1)
step3 = agent.step(utils.send_status(...), depends_on=[step1, step2])

Dynamic Data Access

Use the state object to access runtime data:

# Access nested data
user_id = state.user.id
dataset_config = state.dataset.config.type

# Use in actions
step = agent.step(
    memory.search(query=f"data for user {state.user.id}")
)

Error Handling

Handle errors with conditions and fallback steps:

from erdo.conditions import IsError

main_step = agent.step(llm.message(...))

# Handle errors
main_step.on(
    IsError(),
    utils.send_status(
        message="Analysis failed, please try again",
        status="error"
    )
)

Invoking Agents

Use the invoke() function to execute agents programmatically:

from erdo import invoke

# Invoke an agent
response = invoke(
    "data-question-answerer",
    messages=[{"role": "user", "content": "What were Q4 sales?"}],
    datasets=["sales-2024"],
    parameters={"time_period": "Q4"},
)

if response.success:
    print(response.result)
else:
    print(f"Error: {response.error}")

Invocation Modes

Control how bot actions are executed for testing and development:

Mode	Description	Cost
live	Execute with real API calls	$$$ per run
replay	Cache responses, replay on subsequent runs	$$$ first run, FREE after
manual	Use developer-provided mock responses	FREE always

# Live mode (default) - real API calls
response = invoke("my-agent", messages=[...], mode="live")

# Replay mode - cache after first run (recommended for testing!)
response = invoke("my-agent", messages=[...], mode="replay")

# Replay with refresh - bypass cache, get fresh response
response = invoke("my-agent", messages=[...], mode={"mode": "replay", "refresh": True})

# Manual mode - use mock responses
response = invoke(
    "my-agent",
    messages=[...],
    mode="manual",
    manual_mocks={
        "llm.message": {
            "status": "success",
            "output": {"content": "Mocked response"}
        }
    }
)

Replay Mode Refresh: The refresh parameter forces a fresh API call while staying in replay mode:

• First run: Executes and caches the response
• Subsequent runs: Uses cached response (free!)
• With refresh: Bypasses cache, gets fresh response, updates cache

Perfect for updating cached responses after bot changes without switching modes.

Testing Agents

Write fast, parallel agent tests using agent_test_* functions:

from erdo import invoke
from erdo.test import text_contains

def agent_test_csv_sales():
    """Test CSV sales analysis."""
    response = invoke(
        "data-question-answerer",
        messages=[{"role": "user", "content": "What were Q4 sales?"}],
        datasets=["sales-q4-2024"],
        mode="replay",  # Free after first run!
    )

    assert response.success
    result_text = str(response.result)
    assert text_contains(result_text, "sales", case_sensitive=False)

Run tests in parallel with the CLI:

# Run all tests
erdo agent-test tests/test_my_agent.py

# Verbose output
erdo agent-test tests/test_my_agent.py --verbose

# Refresh cached responses (bypass cache for all replay mode tests)
erdo agent-test tests/test_my_agent.py --refresh

Refresh Flag: The --refresh flag forces all tests using mode="replay" to bypass cache and get fresh responses. Perfect for:

• Updating cached responses after bot changes
• Verifying tests with current LLM behavior
• No code changes needed - just add the flag!

Test Helpers

The erdo.test module provides assertion helpers:

from erdo.test import (
    text_contains,      # Check if text contains substring
    text_equals,        # Check exact match
    text_matches,       # Check regex pattern
    json_path_equals,   # Check JSON path value
    json_path_exists,   # Check if JSON path exists
    has_dataset,        # Check if dataset is present
)

CLI Integration

Deploy and manage your agents using the Erdo CLI:

# Login to your account
erdo login

# Sync your agents to the platform
erdo sync-agent my_agent.py

# Invoke an agent
erdo invoke my-agent --message "Hello!"

# Run agent tests
erdo agent-test tests/test_my_agent.py

Examples

See the examples/ directory for complete examples:

• agent_centric_example.py - Comprehensive agent with multiple steps
• state_example.py - State management and templating
• invoke_example.py - Agent invocation patterns
• agent_test_example.py - Agent testing examples

API Reference

Core Classes

• Agent: Main agent class for creating workflows
• ExecutionMode: Control step execution (iterate, conditional, etc.)
• Prompt: Load and manage prompt templates

Actions

• memory: Store and search memories
• llm: Interact with language models
• bot: Invoke other bots and agents
• codeexec: Execute Python code
• utils: Utility functions (status, notifications, etc.)

Conditions

• Comparison: GreaterThan, LessThan, TextEquals, etc.
• Status: IsSuccess, IsError, IsNull, etc.
• Logical: And, Or, Not

State & Templating

• state: Access runtime parameters and data
• TemplateString: Dynamic string templates with {{variable}} syntax

License

Commercial License - see LICENSE file for details.