Skip to content
Ginkgo

Why Ginkgo

Ginkgo is built on these founding principles:

  1. Scientific workflows belong in Python.
  2. A workflow’s graph should be able to grow as it runs.
  3. Workflows are now written and run by AI agents.

Scientific Workflows Belong In Python

Section titled “Scientific Workflows Belong In Python”

Data scientists and bioinformaticians already write their analyses in Python. Many workflow tools ask them to describe the workflow again in a second language, such as YAML or a custom syntax. The analysis then lives in two languages, and the two can drift apart.

In Ginkgo, a workflow is plain Python functions marked with @flow and @task(). It sits in the same language as the analysis itself.

A Workflow’s Graph Should Be Able To Grow As It Runs

Section titled “A Workflow’s Graph Should Be Able To Grow As It Runs”

Real analyses rarely know their full shape before they start. Take a filtering task that decides how many samples pass a quality check. The next task needs to run once for each sample that passed, and nobody knows that number until the filter has run.

In Ginkgo, a task can look at its inputs and return new tasks. Ginkgo adds them to the graph and runs them. The workflow grows to match the data. Core Concepts shows an example.

Workflows Are Now Written And Run By AI Agents

Section titled “Workflows Are Now Written And Run By AI Agents”

More and more analysis is written, run and debugged by AI coding agents working alongside scientists. A workflow tool should be as easy for an agent to read and operate as it is for a person.

The first two principles already help. Agents write plain Python well. A graph that can grow lets an agent write what the data needs, without working around the tool.

An agent also needs to see what happened in a run. A person can watch a live table in the terminal, but an agent works best from structured facts. So Ginkgo reports runs as JSON. It also records every run in the provenance database, which an agent can query. An agent can then answer “which task made this file?” or “why did this task run again?” from the record, rather than from a guess.

Working With Coding Agents lists the commands that put this into practice.

See Also

Section titled “See Also”