An Alien Mind Needs a Control Plane: help shape LOCO

[ArielSmoliar]

Agent systems are gaining longer horizons, more tools, and more freedom to delegate. Orchestrators decide what agents should do. Independent monitors can estimate whether a trajectory looks suspicious. Infrastructure can contain where code runs.

What turns those inputs into coordinated, fleet-wide action?

LOCO is exploring a specific answer: a provider-independent control plane that decides whether eligible work runs now, later, with less capacity, under review, or not at all.

What exists today

LOCO is an open-source Python scheduling and cost-governance layer. It provides bounded concurrency, backpressure, policy checks, budget enforcement, cost attribution, and load-conscious dispatch across seven agent frameworks and providers.

Its core invariant is that authorization determines eligibility before scheduling begins. Queue depth, age, cost, trust, or adaptive tuning must never turn denied work into eligible work.

What remains open

The monitoring-aware control plane is a research direction. LOCO does not determine whether a model is aligned, replace sandboxing or network isolation, or act as the only kill switch.

The architecture we want to pressure-test is:

1. Orchestrators provide intent and proposed work.
2. Policies define authority and hard boundaries.
3. Independent monitors provide fallible risk findings.
4. LOCO converts those inputs into capacity and intervention decisions.
5. External infrastructure contains execution.
6. Humans retain pause, release, and incident authority.

Questions for collaborators

• What is the minimum event contract a monitor needs before an action executes?
• Which state belongs to an entire trajectory or campaign instead of one agent?
• When should a finding throttle, require approval, pause, quarantine, or terminate?
• How should the fleet behave when a required monitor or containment service disappears?
• What evidence would convince you this reduces blast radius without unacceptable throughput loss?
• Which controls must remain independent of LOCO itself?

Ways to participate

• Challenge the trust boundaries in the threat model.
• Review the monitoring-aware roadmap.
• Take one of the scoped contributor issues.
• Share an agent framework, monitor, incident pattern, or deployment topology that this model fails to cover.

The goal of this discussion is not agreement. It is to find the boundary between a useful fleet control plane and security claims the system cannot support.