@kernel.chat/skill-router

Bayesian skill-rating router for AI agents. Routes tasks to the best specialist using TrueSkill-style ratings that improve with every interaction. Zero LLM calls.

Part of the kernel.chat open-source stack. Used by @kernel.chat/kbot to pick the right specialist for an incoming message without burning a model call on the routing decision itself.

Why this exists

Most multi-agent systems route by hardcoded rules (if intent === 'code' then coder) or by an LLM call on every message (slow and expensive). This package gives you a third option: a Bayesian rating maintained per agent per category. Routing is a constant-time lookup; ratings update from outcomes.

After a few hundred interactions the router's confidence in its assignments is high enough that routing decisions are effectively free.

Install

npm install @kernel.chat/skill-router

Usage

import { SkillRouter, createDefaultRouter } from '@kernel.chat/skill-router'

// Quick start with sensible defaults.
const router = createDefaultRouter()

// Or build your own with explicit config.
const customRouter = new SkillRouter({
  agents: ['coder', 'researcher', 'writer', 'analyst'],
  categories: ['code', 'research', 'writing', 'analysis'],
  // ...config per SkillRouterConfig interface
})

// Categorize an incoming message.
const category = router.categorize('Refactor this function for readability')
// → 'code'

// Route the message to the best agent.
const { agent, confidence, score } = router.route(
  'Refactor this function for readability',
)

// Or route only if confidence is high enough.
const decision = router.routeWithThreshold(message, 0.7)
if (decision) {
  // Dispatch to decision.agent
} else {
  // Fall back to broader routing
}

// After the agent handles the message, record the outcome.
router.recordOutcome(agent, category, 'win')   // 'win' | 'loss' | 'draw'

// Inspect ratings.
const rating = router.getRating('coder', 'code')
const allRatings = router.getAgentRatings('coder')
const topCoders = router.getTopAgents('code', 3)
const systemConfidence = router.getSystemConfidence()

// Persist.
router.save('./router-state.json')
router.load('./router-state.json')

Public API

| Export | Shape | |---|---| | SkillRouter | Main class — categorize, route, routeWithThreshold, recordOutcome, getRating, getAgentRatings, getTopAgents, getSystemConfidence, toJSON/fromJSON, save/load, isDirty | | createDefaultRouter() | Factory returning a SkillRouter pre-seeded with kbot's specialist set | | Rating | Per-agent-per-category rating | | Outcome | 'win' \| 'loss' \| 'draw' | | RouteResult | What route() returns | | SkillRouterConfig | Constructor config |

Status

v1.0.x — production use inside @kernel.chat/kbot; light external test coverage. The router has been driving kbot's specialist selection since early 2026. The public API is stable. External test suite expanded in v1.1.

File issues at github.com/isaacsight/kernel.

Related packages

| Package | Discipline | |---|---| | @kernel.chat/kbot | The agent itself | | @kernel.chat/kbot-orchestrator | Pipeline orchestration | | @kernel.chat/prompt-evolver | Prompt self-optimization from traces | | @kernel.chat/memory-tiers | Three-tier memory | | @kernel.chat/tool-forge | Runtime tool creation |

See docs/agentic-engineering.md for the field map this package sits inside.

License

MIT.