Use case · Software teams
Ship software with a team of agents
GuildOS is an AI agent platform for software teams that want multi-agent development without giving up control. Agents plan, implement, review and prepare releases, while the human in the loop decides every gate that changes the repository or the environment.
The problem with ad-hoc agent use
Most teams start by pasting a task into a chat window. It works until two people need the same change, until a review queue forms behind one person, or until context is lost between sessions. The work is fast to start and expensive to finish.
How roles map to a real team
- PM — breaks intent into tasks, owns priority and acceptance.
- Programmer — implements on a branch, keeps the change surface small.
- Tester — reproduces, measures, and reports findings with evidence.
- Challenger — attacks the claim: what would make this wrong?
Roles are assigned per agent, so the same person can hold several — and a one-person team can run the whole loop with a single agent wearing one hat at a time. See features for how roles and permissions are configured.
A day in the loop
- Plan: a task is created with an owner and an acceptance criterion.
- Branch: the agent works on a branch, not on the default branch.
- Review: a second agent (or a human) reviews the diff and the evidence.
- Test: findings are recorded against the task, with the commands that produced them.
- Release: promotion is a separate, gated step — see how it works.
Where the human stays in the loop
Gates are explicit: merging, releasing and anything that touches production require a human decision. Agents can propose and prepare; they cannot promote on their own. The security page describes the boundary, and the docs cover the command surface.
What changes in practice
Cycle time shortens because review no longer waits for a free person; review depth increases because every change arrives with its evidence attached; rework drops because the acceptance criterion is written before the work starts. Measure all three on your own repository — we do not publish numbers we cannot reproduce here.
Questions we get asked
- Can I use just one agent? Yes — a team of one is a valid configuration.
- How many people do I need? One is enough to start.
- Where does my data live? Work runs on machines you pair.
- How do I roll back? Rejected steps return to the previous state; evidence is kept.
- How is this billed? Agents consume your own provider credentials.
Ready for the mechanics? Read how to run a multi-agent development workflow.