Agent routing

Skill sumitaich1998/jarvisvr/skills/orchestration/agent-routing

An AI agentic operating system for mixed reality on the Meta Quest 3 — your own J.A.R.V.I.S. Multi-agent orchestration, multimodal perception (sight/hearing/gaze), 42 holographic widgets, 20 LLM providers.

Install

npx -y skills add sumitaich1998/jarvisvr --skill agent-routing

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

One thing to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

What its author says it does

Copied from the file, not written here

Route a single sub-task to the best specialist agent (or sub-agent) by matching it against the JarvisVR roster's roles, skill descriptions, and tool sets. Use after task-decomposition for each node, or directly for a one-shot request that clearly belongs to one specialist. Triggers: "who should handle this", route, dispatch, delegate, assign, pick an agent, fan out to sub-agents.

The file declares its own license as MIT. That is the author’s claim about this one file, and it is not the same thing as the license GitHub reports for the repository, which is listed with the other numbers below.

SKILL.md

4.2 KB, as published. Nobody here has run it

Agent Routing

Map one sub-task to exactly one owning specialist (or decide to fan it out to sub-agents). Routing is matching: sub-task intent → role + that role's skill descriptions (which the loader indexed at startup, ~100 tokens/skill).

The roster (docs/ORCHESTRATION.md §3)

Role id	Use it for
`perception-agent`	Seeing the room: identify objects, read/translate signs, describe surroundings, find remembered things.
`research-agent`	Knowledge, web, news, markets.
`productivity-agent`	Timers, reminders, notes, tasks, calendar.
`smart-home-agent`	Lights, climate, locks/cameras, scenes.
`navigation-agent`	Maps, wayfinding, measuring, spatial memory of places.
`media-agent`	Music/video playback, generated imagery, soundscapes.
`communication-agent`	Live translation, captions, drafting messages.
`stage-agent`	Spatial compositor: arrange/anchor/declutter holograms.
`system-agent`	Settings, LLM provider, privacy, diagnostics.

Steps

Extract the action verb + object from the sub-task.
Score candidate roles by matching against each role's skill descriptions and allowed-tools. Prefer the role whose skill description keywords match most directly (this is why descriptions carry trigger words).
Pick the single best owner. Ties → prefer the more specific specialist (e.g. read-and-translate-sign on perception-agent over generic research).
Choose the skill within that agent and attach it as skills:[...] on the plan node and skill on orchestration.agent_status.
Decide on sub-agents (see below) when the sub-task is naturally N parallel items.
Emit agent.thinking with agent_id, role, and skill attributed so the client can highlight who's active.

Routing examples

Sub-task	Role	Skill
"what is this mug?"	`perception-agent`	`identify-object`
"translate this sign"	`perception-agent`	`read-and-translate-sign`
"weather in Tokyo"	`research-agent`	`web-research`
"start a 5-min timer"	`productivity-agent`	`manage-timers`
"dim the living room"	`smart-home-agent`	`control-lighting`
"take me to the kitchen"	`navigation-agent`	`wayfind`
"play some lo-fi"	`media-agent`	`control-media-playback`
"caption this meeting"	`communication-agent`	`caption-conversation`
"tidy these windows"	`stage-agent`	`declutter-space`
"switch to a local model"	`system-agent`	`configure-llm`

Sub-agent fan-out (protocol §9.2 `orchestration.handoff`)

When a sub-task is "do X across N items" (e.g. summarize 3 sources), the owning agent delegates to siblings with dotted ids and emits a handoff:

{
  "plan_id": "uuid-v4",
  "from_agent": "a1",
  "to_agent": "a1.1",
  "to_role": "summarizer",
  "level": 2,
  "subtask": "summarize source 1 of 3",
  "reason": "parallel per-source summarization"
}

Prefer parallel siblings (a1.1, a1.2, a1.3) over deep chains; keep depth ≤ 3.

Edge cases

Multi-role sub-task → it wasn't atomic; send it back to task-decomposition to split.
No confident match → route to research-agent (general knowledge) as a fallback, or clarify-intent if the gap is the user's intent, not the agent.
Rendering-only sub-task (arrange/move/hide) → always stage-agent; other agents describe what, the stage decides where.
Gated capability (camera, lock, spend) → route normally but mark the node so the executing agent requests consent before acting.