---
title: Agentics
slug: /bundles-agentics
---

import Icon from "@site/src/components/icon";
import PartialParams from '@site/docs/_partial-hidden-params.mdx';

The [Agentics](https://github.com/IBM/agentics/) component bundle uses LLMs to transform tabular data. Add or fill columns row-by-row with the [**aMap** component](#amap-component), collapse many rows into one with [**aReduce** component](#areduce-component), or generate synthetic rows with the [**aGenerate** component](#agenerate-component).

Watch it in action with a sentiment analysis example that transforms customer reviews into structured insights:

<div style={{position: 'relative', paddingBottom: '56.25%', height: 0, overflow: 'hidden', maxWidth: '800px', marginBottom: '1rem'}}>
  <iframe
    style={{position: 'absolute', top: 0, left: 0, width: '100%', height: '100%', borderRadius: '8px'}}
    src="https://www.youtube.com/embed/bugrV2pX-9Q"
    title="Agentics Tutorial Video"
    frameborder="0"
    allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share"
    referrerpolicy="strict-origin-when-cross-origin"
    allowfullscreen>
  </iframe>
</div>

## Prerequisites

1. Install the Agentics package in Langflow's virtual environment:

   ```bash
   uv pip install agentics-py
   ```

2. Restart Langflow so the Agentics components are available:

   ```bash
   uv run langflow run
   ```

3. Agentics components require an LLM. Configure your LLM provider API keys as [global variables](/configuration-global-variables) or environment variables.
Supported providers include OpenAI, Anthropic, Google Generative AI, IBM WatsonX, and Ollama.

## aGenerate component

aGenerate generates synthetic data from a schema or from an example DataFrame. Use it for test data, augmentation, or documentation examples.

For example, this schema definition creates the following DataFrame output:

**Schema definition:**
- `customer_name` (str): Full name
- `email` (str): Email address
- `age` (int): Age between 18–80
- `purchase_categories` (str, As List): List of product categories purchased

**Output DataFrame (example, 10 rows generated):**
| customer_name | email | age | purchase_categories |
|---------------|-------|-----|---------------------|
| Sarah Johnson | sarah.j@email.com | 34 | Electronics, Books, Home & Garden |
| Michael Chen | m.chen@email.com | 28 | Sports, Clothing, Electronics |
| ... | ... | ... | ... |

### Parameters

| Name | Type | Description |
|------|------|-------------|
| Language Model | Dropdown | Select the LLM provider and model. Use guided experience. |
| Input DataFrame | Table | Optional. Example DataFrame to learn from; only first 50 rows used. If not provided, Schema is used. |
| Schema | Table | Define columns to generate when no Input DataFrame is provided. See the component's schema definition. |
| Instructions | String | Optional instructions for generation. |
| Number of Rows to Generate | Integer | How many synthetic rows to create. Default: 10. |

## aMap component

aMap transforms each row of input data using natural language instructions and a defined output schema (one row in, one row out).
Use **aMap** for enriching data with LLM-generated columns such as sentiment, categories, and entity extraction.

Rows are processed concurrently; default batch size is 10. Token usage scales with number of rows.

For example, **aMap** keeps each input row and fills in `sentiment`, `confidence`, and `key_topics` with the connected LLM.

**Input DataFrame:**
| review_id | text |
|-----------|------|
| 1 | Great product, fast shipping! |
| 2 | Terrible quality, broke after one use |

**Schema definition:**
- `sentiment` (str): "positive", "negative", or "neutral"
- `confidence` (float): Confidence score 0–1
- `key_topics` (str, As List): Main topics mentioned

**Output DataFrame:**
| review_id | text | sentiment | confidence | key_topics |
|-----------|------|-----------|------------|------------|
| 1 | Great product, fast shipping! | positive | 0.95 | product quality, shipping |
| 2 | Terrible quality, broke after one use | negative | 0.92 | quality, durability |

### Parameters

| Name | Type | Description |
|------|------|-------------|
| Language Model | Dropdown | Select the LLM provider and model. Use guided experience. |
| Input DataFrame | Table | Input DataFrame (list of dicts or DataFrame). Each row is processed independently. |
| Schema | Table | Define the structure and types for generated columns. See the component's schema definition. |
| Instructions | String | Natural language instructions for transforming each row into the output schema. |
| As List | Boolean | If true, generate multiple instances of the schema per row and concatenate. |
| Keep Source Columns | Boolean | If `true`, append new columns to original data; if false, return only generated columns. Ignored if As List is true. Default: `true`. |

## aReduce component

aReduce aggregates all rows in the input DataFrame into a single row following the output schema.
Use **aReduce** for summaries, reports, or consolidated insights.
To aggregate rows into a list, set `As List` to `true` in the component.

All rows are sent in one request; token usage can be high for large DataFrames.
Consider filtering or sampling first.

For example, **aReduce** takes all input rows and produces a single row with LLM-generated aggregates.
It sums revenue into `total_revenue`, identifies the best-selling product in `best_selling_product`, and writes a short `summary` of the sales.

**Input DataFrame:**

| date | product | revenue | units |
|------|---------|---------|-------|
| 2024-01-01 | Widget A | 1200 | 50 |
| 2024-01-02 | Widget B | 800 | 30 |
| 2024-01-03 | Widget A | 1500 | 60 |

**Schema definition:**
- `total_revenue` (float): Sum of all revenue
- `best_selling_product` (str): Product with highest units
- `summary` (str): Natural language summary

**Output DataFrame:**
| total_revenue | best_selling_product | summary |
|---------------|---------------------|---------|
| 3500 | Widget A | Over 3 days, Widget A was the best seller with 110 units, generating $2700 in revenue. |

### Parameters

| Name | Type | Description |
|------|------|-------------|
| Language Model | Dropdown | Select the LLM provider and model. |
| Input DataFrame | Table | Input DataFrame (list of dicts or DataFrame). Required. |
| Schema | Table | Define the structure and types for the aggregated output. See the component's schema definition. |
| As List | Boolean | If true, output is a list of instances of the schema. |
| Instructions | String | Optional instructions for aggregation. If omitted, the LLM infers from field descriptions. |

## Performance and troubleshooting

- **Token usage:** aMap scales with rows; aReduce sends all rows in one call; aGenerate scales with instances. Use smaller batches or sample large datasets to reduce cost.
- **Batch size:** Default 10; max 25. Larger batches improve throughput but increase latency per batch.
- **Errors:** If Agentics is not found, run `uv pip install agentics-py==0.3.1` and restart Langflow. For API/key errors, set the provider’s API key as a global variable or env var. For DataFrame errors, ensure input is a list of dicts or use a DataFrame component output.

## See also

- [Agentics official documentation](https://ibm.github.io/Agentics/)
- [Agentics GitHub repository](https://github.com/IBM/agentics/)
- [Global variables configuration](/configuration-global-variables)

**Publications:**
- Alfio Gliozzo, Naweed Khan, Christodoulos Constantinides, Nandana Mihindukulasooriya, Nahuel Defosse, Gaetano Rossiello, Junkyu Lee. "Transduction is All You Need for Structured Data Workflows." arXiv:2508.15610, 2025. [Link](https://arxiv.org/abs/2508.15610)
- Alfio Massimiliano Gliozzo, Junkyu Lee, Nahuel Defosse. "Agentics 2.0: Logical Transduction Algebra for Agentic Data Workflows." arXiv:2603.04241, 2026. [Link](https://arxiv.org/abs/2603.04241)