@agentskit/observability

See exactly what your agent does — every LLM call, tool execution, and reasoning step — with zero coupling to your agent code.

Tags: ai · agents · llm · agentskit · ai-agents · observability · tracing · opentelemetry · langsmith · logging · monitoring

Why observability

• Debug in minutes, not hours — trace the full ReAct loop: which tools were called, what the LLM received, where it went wrong, all in one place
• Works with your existing tracing stack — LangSmith, OpenTelemetry (OTLP), or a simple console logger; observers are just { name, on(event) } objects
• No coupling, no lock-in — observability attaches to AgentEvent emissions from the runtime; remove it and your agent code is unchanged
• Non-blocking by design — observer errors never surface to the agent; production stability is not at risk

Install

npm install @agentskit/observability

Quick example

import { createRuntime } from '@agentskit/runtime'
import { anthropic } from '@agentskit/adapters'
import { consoleLogger, langsmith } from '@agentskit/observability'

const runtime = createRuntime({
  adapter: anthropic({ apiKey: process.env.ANTHROPIC_API_KEY, model: 'claude-sonnet-4-6' }),
  observers: [
    consoleLogger({ format: 'pretty' }),
    langsmith({ apiKey: process.env.LANGSMITH_API_KEY }),
  ],
})

const result = await runtime.run('Analyze sales data in ./data/sales.csv')
// Every step is now logged and traced automatically

Features

• consoleLogger({ format }) — pretty-print or JSON structured logs for local dev
• langsmith({ apiKey }) — send runs to LangSmith for tracing and evaluation
• OpenTelemetry OTLP exporter — ship traces to any OTEL-compatible backend
• Observer interface: { name: string, on(event: AgentEvent): void } — write custom observers in minutes
• Attaches via observers array on createRuntime — zero changes to agent logic

Ecosystem

Contributors

License

MIT — see LICENSE.

Docs

Full documentation · GitHub