@e-volv/agents

Agent runs as e-volv Observer reads them. Every model call, tool call and agent run becomes a span with the OpenTelemetry GenAI attributes, sent through the @e-volv/logs SDK you already have, so Observer's Agents pages show each run with its tokens, cost, session and prompt version.

npm install @e-volv/agents

import {
  init,
  instrumentOpenAI,
  traceAgent,
  traceTool,
  withSession,
} from '@e-volv/agents';
import OpenAI from 'openai';

init({
  key: process.env.EVOLVE_LOGS_KEY,
  url: 'https://api.e-volv.io/api/public/v1/logs',
  service: 'support-bot',
});
const openai = instrumentOpenAI(new OpenAI());

await withSession({ sessionId: thread.id, userId: user.id }, () =>
  traceAgent('triage', async () => {
    const res = await openai.chat.completions.create({
      model: 'gpt-5-mini',
      messages,
    });
    await traceTool('read_ticket', () => tickets.read(id));
    return res.choices[0].message.content;
  })
);

What it gives you

• init(options): @e-volv/logs options plus content: true to record prompts, completions, tool arguments and results (off by default). e-volv Launch flags are off too, since a Berserker project key has no flags:read; pass flags to turn them on for a key that does.
• traceAgent(name, fn, { agentId?, model?, provider? }): the run.
• traceChat(model, fn, { provider }): one model call; call chat.usage({ input, output, cacheRead, cacheWrite }, responseModel) inside.
• traceTool(name, fn): one tool call.
• withSession({ sessionId, userId }, fn) and withPrompt({ name, version }, fn): tag every span inside.
• instrumentOpenAI(client), instrumentAnthropic(client): wrap the client in place; chat completions, the Responses API and messages, streams included.
• traceMcp(client, serverName): every callTool is a tool span.
• aiMiddleware(): a Vercel AI SDK middleware for wrapLanguageModel.

Already on OpenTelemetry? You do not need this package: point your exporter at https://api.e-volv.io/api/public/v1/logs/otlp/v1/traces with the ingest key as Authorization and Observer reads the GenAI spans directly.