Unattended
Agentic work run unattended, for the hands-on lead. How my agents build software while I'm not watching.
I'm Jens-Christian Fischer. On Unattended, I share real personal and MetaFactory projects: what the agents built, where they went wrong, and what makes their work trustworthy.
The Idea
Agents do the work. I make the decisions.
Writing code is no longer the expensive part. A coding agent can produce a week of changes in a night. What costs now is knowing which of those changes are right, and deciding what should be built in the first place.
Unattended is how I run that in practice. I describe a larger goal as a map of small decisions. Agents pick up the pieces that are decided, build them, review the result and open pull requests while I do something else. They come back to me only when something needs a real decision.
The channel shows the parts that usually stay off camera: the maps, the reviews, the failures, and the rules I added so the same failure doesn't happen twice. Every tool involved is open source.
A Real Example
The first version of this page was built this way. Ranger implemented it from node #2 of a map, Sage reviewed the pull request, and I asked for changes after looking at the preview.
How It Runs
From Goal to Merged Change
Chart
Orienteer, part of Soma, turns a goal too big for one session into a map of decision nodes on GitHub issues. A node closes only when its gate passes.
Walk
Ranger claims decided nodes from the map, works them through a fixed procedure, and files newly discovered work back onto the map.
Review
Sage reviews each pull request through lenses such as code quality, security and architecture before a human looks at it.
Decide
I get the questions only I can answer: scope, taste, trade-offs. Everything else keeps moving without me.
The Work Loop
The Tools Behind the Work
Follow the public repositories to see how each tool is built.
Soma
Keeps an assistant's identity, purpose, memory, skills and working method in one place, and projects them into Claude Code, Codex, Pi.dev and Cursor.
View repository →Ranger
Walks an Orienteer map to completion: claims decided nodes, implements and reviews them through a standard procedure, and escalates only genuine decisions.
View repository →Sage
A code review agent for GitHub and GitLab. Code quality always runs; security, architecture and other lenses join when the change calls for them.
View repository →Assay
A practice for knowing what is true about a system: every claim paired with a comparison that can fail, at the pace agents now ship.
View repository →The Platform Underneath
Cortex
A self-hosted agent bus. Turns an @mention in Discord or on the web into an agent session on your own machine, gated by policy and signed into an audit trail.
View repository →Myelin
The protocol stack that connects agents: signed envelopes, transports, identity and discovery. One schema for every signal.
View repository →Arc
A package manager for agent skills, tools and agents. Search a registry, review capabilities, install with one command, audit what's running.
View repository →The reasoning behind all of it is in the MetaFactory vision.
Get in Touch
Bring Agentic Work to Your Team
Want help putting agents to work in your team? InVisible GmbH offers consulting on agentic software development. Let's talk about your project.
Contact Us