Micro agents that maintain and build your system while you sleep.
Lots of small agents, each with one job, working through the night. In the morning, people review what they did and decide what ships.
A lot of engineering happens when nobody is at the keyboard.
Checking an environment after a deploy, triaging what came in overnight, answering an event as it arrives. A Claude Code session on a laptop stops when the laptop does, and an agent nobody is watching cannot stop halfway to ask a question. So the routine part waits for someone to remember it.
Twenty small agents are easier to trust than one big one.
An agent with one job has a prompt you can read in a minute. It needs only the connectors that job uses, and it can often run on a cheaper model. When it gets something wrong, its transcript shows where, and Revert puts the last good version back.
What every Reactor feature has to keep true.
People set the direction.
A person decides what each agent is for and reviews what it hands back. Even a supervisor acts as the person who turned it on, and that person's own change takes control back.
You can see everything an agent does.
Every agent is a card on a canvas, every run has a transcript, and every change to an agent is a version you can revert.
You can see what the work costs.
Every run records its cost and tokens, and a supervisor's default mission is to make its workspace cheaper without doing less.
Triggers, memory, workspaces and supervisors are built. Next is a fleet that looks after a whole codebase.
Triggers and memory
Agents that start themselves and remember the last run.
Workspaces and groups
A fleet you can see, sorted by team and by job.
Supervisors
An agent that tunes the others in its workspace.
Whole codebases
Fleets that keep a system current and working, with people steering.