Guardrails for Autonomous Agents
Fanning an agent out across many workers is powerful and risky. Arbiter provides configurable controls for stopping work, limiting depth, setting budgets, and requiring approval, while interrupted side-effecting work remains visible for review.
Letting an agent run on its own is useful right up until it isn’t. The same autonomy that clears a backlog overnight can also spawn a swarm of workers, chew through a budget, or recurse into itself and never stop. The power and the terror are the same feature. What makes autonomy usable is not trusting the agent to behave. It is the set of controls that stay in your hands while it works. Arbiter provides those controls so operators can configure a posture that matches the work.
One tap stops everything
When something looks wrong, you do not want to reason about which worker to kill. You want to halt the whole thing, now, and sort it out after. Arbiter provides a stop-all control for active work. Use it to halt the operation, inspect the state, and decide the next step from a known position.
Caps that a runaway job cannot argue with
An agent asked to parallelize can decide the answer is more agents, and those agents can decide the same thing. Arbiter supports limits on fan-out, recursion depth, concurrency, and budget. Configure the limits you need, then verify them against the workload you intend to run.
The budget is a wall, not a warning
Cost is the failure that hurts quietly, because you do not see it until the bill does. Arbiter records budget information and can enforce a configured budget ceiling. A ceiling is a policy choice, not an unavoidable default, so set it before trusting it with a paid workload.
Fan-out asks first
Going wide is a big decision, so it is one you make, not one that happens to you. Fan-out approval can be required above a configured threshold. That makes expansion deliberate when your operating policy calls for it.
Nothing gets to run forever in the dark
Long-running autonomy has a specific nightmare: a job that quietly keeps going with no one watching. Arbiter records interrupted runs and keeps side-effecting work suspended for operator review instead of blindly replaying it. Host-level restart behavior depends on the service or watchdog you configure.
Guardrails are the product
An orchestration layer that only dispatches is easy. One you would actually trust with autonomy is the hard part, and it is the part that has to hold when you are not looking. Stop controls, limits, budget policy, fan-out approval, and deliberate recovery are tools for keeping supervision visible.
Arbiter is a proprietary working alpha in public binary release preparation. Follow the public release channel or RSS feed for updates.