Agentic ecology

Agentic ecology provides AI-driven tools to assist researchers and practitioners processing, analyzing, and annotating their data. It currently focuses on passive acoustic monitoring (PAM) datasets (bioacoustics), with more capabilities to be introduced over time.

Using an agentic development platform such as Google Antigravity, you can generate custom python scripts, build interactive web interfaces to listen to and search your recordings, and train classifier models—all by describing what you want to do in plain English.

Why?

Setting up an entire agentic development framework for you to build a data processing app sounds like a whole lot of trouble when we could simply develop the app itself and distribute it to you, doesn't it?

The reason for that is empowerment: we want you to have the ability to build software tools that address your needs. We recognize that doing so requires its own skillset, but we firmly believe that agentic coding, paired with the right set of agent skills, can eliminate that barrier.

Think of this project not as a tool, but as a tool-building tool.

---

🚀 Getting started

Warning

Unless you have extensive experience with agentic software engineering, we recommend that you set up your agentic coding environment so that the agent asks your permission before performing any operation on your machine on your behalf or to interact with files outside of the workspace's root directory. Vet every command the agent intends to run on your behalf and carefully inspect the code it requests to execute, as agent mistakes can happen. Avoid letting the agent perform operations on your data without backing it up and taking other relevant precautions.

1. Verify and install prerequisites

Verify that the following prerequisites are installed:

• node >=22.20.0
• npx
• git
• uv

If you need help on this, prompt your agent:

Check if node >= 22.20.0, npx, git, and uv are installed. For any that are missing or below the required version, detect my OS and install/upgrade them for me.

2. Install the agentic-ecology-init skill globally

Prompt your agent:

Use npx skills to install the agentic-ecology-init skill from google-deepmind/agentic_ecology globally. Make sure it is globally discoverable for you.

Or if you are comfortable with the command-line, run:

npx skills add google-deepmind/agentic_ecology --skill agentic-ecology-init -g

3. Initialize a new project directory

Open your agentic development platform and prompt your agent:

Use the agentic-ecology-init skill to initialize an agentic ecology project in <my_project_directory>.

---

🛠️ Demo workflow

To help you get a feel of what's possible with agentic ecology, let's simulate a task in which we are trying to bootstrap the annotation of a passive acoustic monitoring dataset for a targeted bird species. We will work with the Powdermill dataset published by Chronister et al. (2022) and search for Hooded Warbler songs.

Ask the agent

I would like to analyze the contents of the Powdermill PAM dataset. The data is hosted here: https://zenodo.org/records/4656848/files/mp3_Files.zip. Download and extract the files into data/powdermill, then help me get started.

What the agent does

The agent:

1. Downloads and unzips the Powdermill audio data in data/powdermill.
2. Identifies all audio files in that directory.
3. Runs the Perch 2.0 model over them to build an audio index.
4. Creates a vector database in the databases/ folder of the project to store everything. (This step may take some time, especially if executing on CPU. Expect around 15 minutes.)
5. Builds a web application for you to browse, search, and annotate the database.
6. Presents you with instructions on how to access and use the web application.

Try out the web app

Use the agent instructions to access the web app, and try searching for Hooded Warbler songs in the database. You can use a recording from Xeno-Canto to get started by entering xc565524 into the query URI bar.

Ask the agent to make changes to the web app

I would like to be able to filter the audio windows by the recording from which they are taken. Add a UI element for that.

The agent will autonomously figure how to modify the existing code to accomplish that, restart the backend server, and prompt you to reload the webpage.

---

💡 Key takeaway

This repository is very minimal: it contains reference dependency configurations and workspace guidelines (stored in skills/agentic-ecology-init/assets) along with a collection of modular capability instructions (stored in skills), all of which are human-readable.

The agent instructions and skills are nothing more than a shortcut that reliably sends the agent in the right direction: they were themselves constructed by prompting the agent to achieve a particular outcome, interactively solving the problem with it, and asking it to write instructions for its future self to arrive at to the solution right away. When the agent made a mistake, it was asked to reflect on it and to amend the instructions and skills so that it doesn't make the same mistake again in the future.

Any ecological analysis capability that the agent currently has could just as well be achieved without any skill or pre-written instructions by working with it iteratively and interactively. In fact, this is exactly how we intend to expand the agent capabilities in this project!

This means that building the right tool for your needs is within your reach: don't hesitate to state your needs; to question the agent; to ask it to clarify, self-correct its mistakes, and amend its instructions and skills; to nudge it in the right direction if it starts veering down the wrong path.

This is not an officially supported Google product.