Skip to content

sumitaich1998/jarvisvr / skills/orchestration/agent-routing

agent-routing

Verified

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.

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 idUse it for
perception-agentSeeing the room: identify objects, read/translate signs, describe surroundings, find remembered things.
research-agentKnowledge, web, news, markets.
productivity-agentTimers, reminders, notes, tasks, calendar.
smart-home-agentLights, climate, locks/cameras, scenes.
navigation-agentMaps, wayfinding, measuring, spatial memory of places.
media-agentMusic/video playback, generated imagery, soundscapes.
communication-agentLive translation, captions, drafting messages.
stage-agentSpatial compositor: arrange/anchor/declutter holograms.
system-agentSettings, LLM provider, privacy, diagnostics.

Steps

  1. Extract the action verb + object from the sub-task.
  2. 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).
  3. Pick the single best owner. Ties → prefer the more specific specialist (e.g. read-and-translate-sign on perception-agent over generic research).
  4. Choose the skill within that agent and attach it as skills:[...] on the plan node and skill on orchestration.agent_status.
  5. Decide on sub-agents (see below) when the sub-task is naturally N parallel items.
  6. Emit agent.thinking with agent_id, role, and skill attributed so the client can highlight who's active.

Routing examples

Sub-taskRoleSkill
"what is this mug?"perception-agentidentify-object
"translate this sign"perception-agentread-and-translate-sign
"weather in Tokyo"research-agentweb-research
"start a 5-min timer"productivity-agentmanage-timers
"dim the living room"smart-home-agentcontrol-lighting
"take me to the kitchen"navigation-agentwayfind
"play some lo-fi"media-agentcontrol-media-playback
"caption this meeting"communication-agentcaption-conversation
"tidy these windows"stage-agentdeclutter-space
"switch to a local model"system-agentconfigure-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.
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall