Sovereign Fleet: a sandboxed task queue that runs goose unattended, with measured routing

[cipherfoxie]

Sovereign Fleet runs coding-agent CLIs unattended on your own machine: you queue a task, it runs in its own git worktree inside a Landlock + seccomp sandbox, the repository's own tests decide whether it lands, and a router learns which harness and model pair finishes a kind of task fastest.

Why it is here: goose is the default harness, because it measured fastest. We ran goose, Aider, Pi and opencode 192 times on the same local Qwen3.6 35B and the same tasks; goose needed 1.6 minutes per correct result, Aider 3.0, Pi 5.0, opencode 6.5. Method and raw data: https://sovgrid.org/blog/goose-aider-pi-opencode-local-qwen-192-runs/

What it gives goose:

• an MCP extension (12 tools) so a goose session can queue fleet tasks, watch them, cancel them and read their receipts;
• three recipes in goose's one directory per recipe layout (A/B evaluation of a change to the agent's rules, a self-improvement loop over failed tasks, a benchmark runner); they call an evaluation script you bring, the README of each says what it gets and must print;
• unattended goose run with its own stop rules (--no-session --no-profile, max turns, tool repetition limit) and a retry when goose crashes at start.

What it does not do yet: Linux only (Landlock), no macOS sandbox; alpha (0.6.1); the router has learned mostly from one local model so far.

pip install sovereign-fleet · https://github.com/Sovgrid/sovereign-fleet · project page https://sovgrid.org/fleet/

Feedback on the recipes is the most useful thing for us: they are our first attempt at goose's recipe format.