Orkestryn

AI agents can each make a reasonable decision and still cause problems when they change the same service. A deployment agent might request capacity while another agent changes the budget. Recovery might stop a failed deployment but also remove resources that a customer still needs.

I built Orkestryn to explore a practical question: how can several AI operators coordinate changes to one service while respecting shared constraints and preserving legitimate work during recovery?

Orkestryn brings planning, policy governance and recovery into an authenticated operations dashboard. Three agent roles use NVIDIA Nemotron through Nebius Token Factory to propose scenarios, order changes and recommend recovery. A separate controller checks dependencies, capacity, ownership, versions and approval scope. A model response is a proposal; it does not grant permission to execute.

Users inspect the exact actions and arguments before approving an operation. The system records execution receipts in SQLite, reconciles interrupted work and monitors service health. Recovery is scoped to eligible effects rather than resetting the whole environment. Manager and customer-owner handoff reviews are separate, and releasing a finished job requires its own approval.

I also built Safety Lab to investigate the oversight side of the problem. It compares controls across 48 authored cases in six families and supports separately reviewed live candidate-generation and monitor requests. An independent simulated environment scores what the proposed actions actually do. Results distinguish harm prevention, safe task completion, false positives and coverage, with evidence available for export and replay.

The application uses Python, SQLite and a JavaScript browser interface. Nebius hosts NVIDIA inference, so it runs on my laptop and desktop without a local NVIDIA GPU. The downloadable judge workspace includes a free scripted walkthrough with real local worker processes and HTTP effects. Live model requests are explicitly reviewed, counted separately and never retried automatically.

The hardest parts were handling uncertainty honestly: an accepted request is not necessarily a completed effect, a lost response does not justify repeating an operation, and a model's explanation is not proof that its plan is safe. Early live Safety Lab replies sometimes failed the tool or argument contract. Improving the request schema and retaining response diagnostics helped make those failures inspectable without accepting unsupported actions.

I verified 945 offline tests on Linux, covering policy checks, approvals, execution, recovery, restart handling and evidence. Live monitor experiments also showed why evaluation matters: one authored attack was detected, while a safe generated plan received a false positive. These are small experiments, not a claim of general monitoring reliability.

This project taught me that useful agent autonomy depends on explicit authority, durable state and clear evidence—not just a convincing answer. Next, I want to expand the scenarios, compare independently chosen monitors and study how inspection budgets affect the trade-off between catching harmful actions and allowing legitimate work.

Orkestryn is a working local prototype. Its capacity and customer-resource scenarios are fixtures, and its measurements do not establish production safety or cloud performance.