At our software company, we wanted autonomous agents that could own real recurring jobs for the team. Not a coding assistant or a personal agent on my laptop, a hosted agent with its own job, like gathering product feedback and maintaining the roadmap or watching competitors and reporting insights back to the team. But we had a hard time finding a sensible place for them to live and a reliable way for them to work with the rest of the team. The options felt like too much or too little.
A routine tied to one vendor and one person, a personal assistant pressed into server duty, or an entire multi agent platform when we just wanted to add one agent with one job at a time. And most agents still work in a bubble. They do something, produce an output for a person, and then disappear. The team can’t see what they’re planning, what they’ve learned, or where they’re stuck.
So we built OpenRoutines, an open source framework for running simple, secure, and durable autonomous agents that act like teammates. The core idea is simple. The repo is the agent. Its routines and knowledge live in Git, and the whole thing deploys as a container.
Let me show you how it works. After you install OpenRoutines, the new command creates an agent and its Git repository. Its job description, routines, skills configuration, and encrypted credentials live here. Persisted knowledge is stored in a special branch.
You can review it like code, test it locally, and deploy it anywhere that runs a Docker container. Here’s a real example, MarketScout. It finds conversations an audience is having, watches competitors, and produces insights for the team. Each part of the job is a routine stored in a markdown file.
This one is called Topic Radar. Its front matter says when it runs and exactly which skills, credentials, and tools it can use.
The body describes the work it’s doing, producing briefs for the team. There’s no orchestration graph or hidden configuration. The agent’s behavior and permissions are right here in files you can review and change. We can check and rehearse our routines locally using the same container we’ll use in production, but against data fixtures with no credentials or external actions. And when we’re ready to deploy, we just need to find a host that deploys Docker. Here, I’m deploying to a VPS by pushing my changes to the main branch. MarketScout runs a handful of routines like this.
They all feed into OpenRoutines’ built in teamwork system, which tracks what happened, what still needs doing, and what’s worth keeping as future context. Here, a reporting routine brings that into Steady, where we can work with the agent alongside the rest of the team. It could also report to Slack, Discord, email, or wherever your team can pick it up. Everyone can see what it did, what it plans to do next, and where it needs help. OpenRoutines is open source with templates and plugins to help you get started.
Or you can ask your coding assistant to build an agent with you. We baked in guidance for that too. So spin up an agent, give it a job, turn that job into a few markdown routines, and put it to work for your team. You can find it all at openroutines.dev.
