SwarmOS & Goose AI Orchestration

Give one AI agent a task and it can go off the rails. Give ten agents a task and coordinate them badly, and now you’ve got ten ways to go off the rails at once. This project builds an orchestration engine from scratch — teams of AI agents run from declarative configuration files, each role routed to an appropriate model across providers.

I also contributed a parallel execution concurrency engine in Rust to Goose, an open-source agentic coding framework, handling independent subtasks and review loops gated on a critic’s verdict.

Enforcing Safety in Code

Autonomy is genuinely useful; autonomy nobody signed off on is a liability. I actively removed a built feature that would let a model invent its own workflow on the fly, replacing it with workflows a person approves first.

Safety and cost controls are enforced strictly in code:

• Ownership rules dictating what each agent may write.
• Spending and runtime circuit breakers.
• Restricted tool access and policy filters.
• Full event logging backed by a comprehensive 300+ test suite.

Driving a Real Robot

The same orchestration kernel, unmodified, was bridged to ROS 2 as a stdio MCP server. I gave the deterministic mock and the live rclpy backend an identical tool surface (like 8-sector lidar summarization rather than raw scan floats), so behavior developed offline transfers to hardware unchanged.

Safety envelopes are enforced in the backend, below any planner:

• Linear and angular speeds are clamped to platform limits, regardless of what the model requests.
• A front-cone obstacle guard converts would-be collisions into refused actions.

Tested live against ROS 2 Jazzy and Gazebo, the system completed a 4.78 m patrol with closed-loop drives landing within ~2 mm of the request. Every safety intervention surfaced correctly in the run report.