caribeedu/roboratory

A factory for AI agents. You build production lines out of machines, belts, signals and gates. Agents work the machines, and you watch the whole floor in pixel art.
Roboratory is for work that has to be repeatable and auditable: turning tickets into pull requests, answering incidents, shipping to production. Each line is a deterministic workflow defined in files and versioned in git. The agents only produce what goes inside each step, and every step's output has to pass a contract before it moves on.
## Why
Running several AI agents on real work today means either free-form chat between agents, which is hard to predict, or workflow engines that hide the agents behind logs. Roboratory treats agent work like a factory floor:
Deterministic lines. Routing, checks and gates are defined in files, so the same input takes the same path every time.
Contracts at every step. An output that does not match its JSON Schema runs the step again; when attempts run out, the machine catches fire and the line stops behind it.
Humans hold the gates that matter. Releasing a gate or applying a change is never automatic, and a machine that wants to run a command outside its list waits for you to allow it. Managers answer questions and propose, and you decide.
Everything visible. If something is running, waiting or burning, it shows on the floor, on the minimap and in the log.
Git is the source of truth. The factory is a repository. Changing it means a reviewed diff, never improvisation at runtime.
Local. No telemetry, open file formats, and your own agent CLIs and subscriptions.
## Build any flow you want
Roboratory does not come with one fixed pipeline: you design the flow, and it runs exactly as designed. A line is a set of pieces you can connect however your process needs:
A sequence of machines, with signals and gates wherever a check or a decision belongs.
Branches: route each box on a field of its contract (type: bug one way, type: question the other).
Joins: a piece can wait for several belts to arrive before it runs.
Loops with a limit: send work back for rework (a failing test, a reviewer's findings), with a lap limit so a line cannot spin forever.
Many lines from one model: one incident line per service, one release line per repository, all kept in sync.
Your own sectors: group lines by team or purpose, each with its own manager and policy.
You can write these as plain YAML in the factory repository, or build them with clicks on the Floor in build mode, which checks the line before saving.
## One run, start to finish
A run starts when a work order enters at an intake, and every run follows the same path: each step produces a box, the box is checked before it moves on, and the last box lands on the pallet. No agent decides that path, each machine starts with a fresh context, and every hop is recorded in the run's logs, which the Ledger shows.
### The pieces
| Piece | Concept | What it is |
| Machine | One agent | Runs one headless agent (Claude Code) with a role and a contract. The step ends when the process exits. |
| Box | Outcome | The output of a machine, validated against its contract. It is the unit of progress that travels along the line. |