TAgent: Build Powerful AI Agents, Not Boilerplate

A developer-first framework for crafting everything from simple AI assistants to complex, multi-agent workflows with elegance and ease.

┌─────────────────────────────────────────────────────────────────────────────────┐
│                           TAgent Architecture                                   │
├─────────────────────────────────────────────────────────────────────────────────┤
│                                                                                 │
│  Simple Agent (run_agent)           Multi-Step Pipeline System                  │
│  ┌─────────────────────┐            ┌─────────────────────────────────────────┐ │
│  │                     │            │                                         │ │
│  │   Goal → Agent      │            │    Step 1 → Step 2 → Step 3             │ │
│  │      ↓              │            │      ↓       ↓       ↓                  │ │
│  │   [Tools]           │     →      │   [Tools]  [Tools]  [Tools]             │ │
│  │      ↓              │            │      ↓       ↓       ↓                  │ │
│  │    Result           │            │   Conditional Execution & Dependencies  │ │
│  │                     │            │      ↓       ↓       ↓                  │ │
│  └─────────────────────┘            │   Structured Outputs & State Flow       │ │
│                                     │                                         │ │
│  Perfect for:                       │  Perfect for:                           │ │
│  • Quick tasks                      │  • Complex workflows                    │ │
│  • Single-step operations           │  • Multi-step processes                 │ │
│  • Simple automations               │  • Conditional logic                    │ │
│                                     │  • Parallel execution                   │ │
│                                     │                                         │ │
│                                     └─────────────────────────────────────────┘ │
└─────────────────────────────────────────────────────────────────────────────────┘

Why TAgent?

• Focus on Your Logic: Stop wrestling with complex frameworks. Write standard Python functions, and TAgent's ToolExecutor intelligently adapts to them.
• Build, Don't Just Prompt: Move beyond simple prompting. Create robust, stateful agents that can plan, execute, and learn from their actions.
• Scale with Confidence: Start with a single agent for a simple task. As your needs grow, scale up to a multi-step, parallelized workflow with the powerful Pipelines engine.
• Universal Compatibility: TAgent is model-agnostic. By using structured JSON outputs instead of proprietary function-calling, it works with hundreds of LLMs out-of-the-box.

Features at a Glance

• 🧠 Task-Based Agents: Predictable, state-driven agents that plan and execute to achieve goals.
• 🛠️ Developer-First Tools: Your Python functions are first-class citizens. No boilerplate required.
• 🚀 Powerful Pipeline Engine: Orchestrate complex, multi-step workflows with dependencies, parallelism, and advanced data flow control.
• 🔗 Smart Data Injection: Use read_data to automatically inject outputs from previous steps as tool parameters, enabling sophisticated prompt enhancement.
• 🌐 Model Agnostic: Compatible with any LLM provider, including OpenAI, Google, Anthropic, and more via LiteLLM.
• 🔒 Structured & Reliable: Enforces structured outputs for predictable, type-safe results using Pydantic.
• 🤖 Simple & Scalable: Start with a few lines of code and grow to production-grade automations.

Quick Look

Simple Agent Tasks

See how easy it is to create an agent with custom tools for simple tasks.

from tagent import run_agent

# 1. Give your agent a goal
goal = "What is the current stock price for NVDA and should I buy it?"

# 2. Give it a tool (a simple Python function)
def get_stock_price(symbol: str):
    """A tool to get the latest stock price for a stock symbol."""
    print(f"--- Getting price for {symbol} ---")
    # (Your logic to call a real stock API would go here)
    if symbol == "NVDA":
        return "stock_price", {"symbol": "NVDA", "price": 950.00}
    return "stock_price", {"symbol": symbol, "price": "unknown"}

# 3. Run the agent
result = run_agent(
    goal=goal,
    tools=[get_stock_price],
    model="gpt-4o-mini"
)

print(result.final_output)

Multi-Step Workflows with Pipelines

For more complex scenarios, use TAgent's Pipeline system to build sophisticated workflows:

from tagent.pipeline import PipelineBuilder
from tagent.pipeline.conditions import IsGreaterThan, IsLessThan

from tagent.pipeline.executor import PipelineExecutor, PipelineExecutorConfig
from tagent.config import TAgentConfig

from pydantic import BaseModel, Field


# Define structured outputs
class SentimentAnalysis(BaseModel):
    score: float = Field(description="Sentiment score from 0-10")
    category: str = Field(description="positive, negative, or neutral")

class EmailDraft(BaseModel):
    subject: str = Field(description="Email subject line")
    body: str = Field(description="Email body content")

# Build a customer feedback pipeline
pipeline = PipelineBuilder(
    name="customer_feedback_pipeline",
    description="Process customer feedback and respond appropriately"
).step(
    name="analyze_sentiment",
    goal="Analyze the sentiment of customer feedback",
    output_schema=SentimentAnalysis
).step(
    name="send_thank_you",
    goal="Draft a thank you email for positive feedback",
    depends_on=["analyze_sentiment"],
    condition=IsGreaterThan("analyze_sentiment.score", 7.0),
    output_schema=EmailDraft
).step(
    name="escalate_complaint",
    goal="Create escalation ticket for negative feedback",
    depends_on=["analyze_sentiment"],
    condition=IsLessThan("analyze_sentiment.score", 4.0)
).build()

executor_config = PipelineExecutorConfig(
    max_concurrent_steps=3,
    enable_persistence=False
)

config = TAgentConfig(model="gpt-4o-mini")

executor = PipelineExecutor(pipeline, config, executor_config)
result = await executor.execute()

Installation

Get started in seconds. Install the core agent, or include optional extras like the Pipeline engine.

# Install the core agent
pip install tagent

# Install with the Pipeline engine
pip install "tagent[pipeline]"

# Install everything
pip install "tagent[all]"

Dive Deeper

This README is just a glimpse of what TAgent can do. For detailed guides, tutorials, and API references, please visit our full documentation.

➡️ Read the Full TAgent Documentation

Our documentation covers:

• Core Agent: Getting started, creating tools, and configuration.
• Pipelines: Building complex, multi-step workflows with dependencies, parallel execution, and advanced data flow features like read_data.
• API References and more.

Contributing

We welcome contributions! Please see our contributing guide for details on how to get started.

License

This project is licensed under the MIT License. See the LICENSE file for details.