agent-relay - Letting AI Coding Assistants Work as a Team
Business Goal
Developers increasingly use AI assistants to write code, and each assistant has different strengths. agent-relay lets a developer stay in the one assistant they are already talking to and have it hand work to the others, such as a second opinion or a review, then report back clearly whether the work actually succeeded.
What Needed to Happen
- ▸ Work with the four most common AI coding assistants through one simple, consistent command
- ▸ Give a clear answer at the end of every task (done, partly done, blocked or failed) instead of guesswork
- ▸ Let assistants use it on their own, without the developer copying text between windows
- ▸ Pass a task to a different assistant when one runs out of capacity halfway through
- ▸ Show progress live on a simple local dashboard, with nothing sent to the cloud
- ▸ Install in one step, detecting which assistants are already on the computer
Delivery Challenges
- - Telling a task that really succeeded from one that only appeared to finish
- - Picking up an earlier conversation with an assistant without it stopping to ask questions
- - Handing over the full picture (goal, files touched, unfinished changes) from one assistant to another
- - Cleaning up after each task so nothing keeps running in the background
- - Keeping everything working as each assistant's tools change over time
What We Delivered
- - Wrapped each assistant in a small command that behaves the same way and returns the same kind of result
- - Added a standard interface that lets assistants start, check, continue, hand over and stop tasks on their own
- - Built a handover that packages the goal, touched files and unfinished work for the next assistant
- - Added automatic fallback to another model when one hits its usage limit
- - Built a local dashboard to watch tasks live and spot the ones that need attention
- - Wrote plugins and instructions so each assistant learns how to delegate
- - Published it as a free, open-source package with an installer, a health check and automated tests
Released free for anyone to use (version 1.3.2, MIT licence) and working with four AI coding assistants. We use it every day to build FitSchool: one assistant plans and checks the work while others build parts of it in parallel.