You write the workflow. It hums.

You define the agents and their stages in YAML. Goobers runs them and gates the result before it opens a pull request.

workflows/implementation.yaml
1name: implementation
2on:
3 trigger: backlog-item
4stages:
5 plan:
6 goober: planner
7 implement:
8 goober: implementer
9 test:
10 run: goobers test
11 gate:
12 kind: agentic
13 reviewer: qa-goober
14next: open-pr
↻ repass trigger plan implement test gate pr
A goober, the Goobers mascot, standing beside the landed pull request
A goober

01 · goober

This is a goober.

An agent, defined in YAML.

A goober

02 · workflow

This is a workflow.

The stages it runs, and the gates that check them.

trigger implement gate pr
A goober A goober A goober

03 · gaggle

This is a gaggle.

A group of workflows, working together.

This is Goobers. Managed, concurrent workflows for agentic velocity.

01

Change the config, or fork the code

Config you own

A gaggle's shape is plain YAML and markdown in a config repo you own: goobers, instructions, workflows. Change it and restart to pick up the new state. Canonical examples ship in the binary too: goobers examples show implementation prints the YAML directly, no checkout required. Full reference: config-examples/.

config-examples/
├── manifest.yaml
└── gaggles/
    └── acme-web/
        ├── gaggle.yaml
        ├── workflows/
        │   └── implementation.yaml
        └── goobers/
            └── implementer/
                ├── goober.yaml
                └── instructions.md

Code you can fork

The rest is MIT-licensed: scheduler, runner, and journal in Go; portal in React, embedded in the same binary. Fork it, or leave it alone and change only the config. Harness support: two adapters, Copilot CLI and Claude Code, behind the same registry. Config reference: the docs.

02

Real work, from the record

Goobers builds itself: the project's own instance runs its workflows against this repo and merges its own PRs under gate policy. Issue #1451 asked for a typed notification-output sink with durable delivery receipts. The implementation workflow claimed it, repassed three times against both an agentic reviewer and local CI, and opened the PR below.

PR #2569 · MERGED — +1743 / −3 across 23 files, 1h 52m

03

What you bring

You run Goobers yourself — there is no managed service. Bring a laptop or server (Linux or macOS first; Windows has its own verified path), sign in a coding agent of your choice (Copilot CLI or Claude Code), and reference your GitHub or Azure DevOps credentials by name, never inline. What it actually runs against is a repository and a backlog: issues or work items you've already labeled ready, the same queue your team triages today.

The loop runs unattended; deciding what's in it doesn't. You still pick what lands on the backlog, what the product needs next, where things head — Goobers turns that backlog into pull requests, it doesn't invent the backlog for you.

goobers

Your gaggle is ready.

Build it and point it at a backlog. Read the journal in the morning. It hums.