CoDA: Agentic Systems for Collaborative Data Visualization

CoDA: Agentic Systems for Collaborative Data Visualization
Zichen Chen, Jiefeng Chen, Sercan O. Arik, Misha Sra, Tomas Pfister, Jinsung Yoon
International Conference on Learning Representations (ICLR 2026)

CoDA (Collaborative Data-visualization Agents) is a multi-agent framework that turns natural language queries into publication-quality visualizations. Instead of treating visualization as a monolithic code-generation problem, CoDA distributes the work across 8 specialized LLM agents in four sequential phases — Understanding, Planning, Generation, and Self-Reflection — that iteratively refine the output until explicit quality thresholds are met.

Results

Architecture

CoDA employs 8 specialized LLM agents across 4 phases. A global TODO list is generated by QueryAnalyzer at the start and is propagated through every subsequent agent; VisualEvaluator uses image-level analysis to verify completion before the workflow terminates.

Gallery

CoDA consistently produces complete, accurate visualizations scoring 90–95/100, while baselines frequently produce broken outputs scoring 0–45/100.

Selected outputs

Self-reflection recovery

CoDA's self-reflection loop detected layout and labeling errors across 3 iterations, then automatically recovered at iteration 4 — demonstrating the robustness of quality-driven refinement.

Installation

Requirements: Python ≥ 3.8 and API credentials for at least one LLM provider.

git clone https://github.com/google-research/agentic-visualization.git
cd agentic-visualization
pip install -r requirements.txt

Configure your API keys:

cp .env.example .env
# Fill in OPENAI_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, etc.

Usage

Simple API

import coda

result = coda.plot(
    query="Create a sunburst chart showing browser market share by version.",
    data="path/to/data.csv",   # CSV, JSON, Excel, SQLite, or a directory
    output="result.png",
)
print(f"Quality: {result.quality_score:.2f}  —  saved to {result.output_file}")

Programmatic Workflow

from coda import Workflow
from coda.workflow.orchestrator import WorkflowConfiguration

cfg = WorkflowConfiguration(
    quality_threshold=0.9,
    max_iterations=3,
    enable_search_agent=True,
    output_directory="outputs/",
)

wf = Workflow(config=cfg, query_id="q001")
result = wf.execute_workflow(
    query="Plot monthly active users with a 90-day rolling average.",
    data_input="data/mau.csv",
    workflow_context={"original_data_path": "data/mau.csv"},
)

print(f"Success: {result.final_success}  |  Quality: {result.final_quality_score:.2f}")
# Output image: outputs/<model>_query_q001/query_1/final_result.png

Benchmarks

1 — Download benchmark data

bash prepare_benchmark.sh
# Downloads MatPlotAgent benchmark into benchmark_data/

2 — Run generation

python run_benchmark.py \
    --start 1 --end 100 \
    --model-name        openai/gpt-4o \
    --search-model-name openai/gpt-4o-mini

Results are saved to benchmark_outputs/<model>_query_<id>/.

3 — Run evaluation

python run_evaluation.py \
    --results-dir   benchmark_outputs \
    --start 1 --end 100 \
    --model-name         openai/gpt-4o \
    --vision-model-name  openai/gpt-4o \
    --processes 10

Model tags follow the LiteLLM naming convention (provider/model-name). Any provider supported by LiteLLM — OpenAI, Anthropic, Google Gemini, Groq, etc. — works out of the box.

Citation

@inproceedings{chen2026coda,
  title     = {{CoDA}: Agentic Systems for Collaborative Data Visualization},
  author    = {Chen, Zichen and Chen, Jiefeng and Arik, Sercan O. and
               Sra, Misha and Pfister, Tomas and Yoon, Jinsung},
  booktitle = {International Conference on Learning Representations},
  year      = {2026},
  url       = {https://coda-agent.github.io/CoDA/}
}

License

Apache 2.0 License — see LICENSE for details.

Disclaimer

This is not an officially supported Google product. This project is not eligible for the Google Open Source Software Vulnerability Rewards Program.

This project is intended for demonstration purposes only. It is not intended for use in a production environment.