AI Agent Observability: Add Logging, Tracing, and Monitoring to Your Python Agents

Every article about AI agent observability tells you to buy a platform. Dynatrace. Arize. Galileo. They're great tools — if you have an enterprise budget and a dedicated observability team.

If you're a developer deploying autonomous agents in Python, you need observability that ships with your agent, not a separate SaaS subscription. Here's how to build it.

What "observability" means for an AI agent

Traditional service	AI agent
Request latency	Agent run duration
Error rate	Tool call failures, LLM refusals
Throughput	Runs per day, tokens consumed
Log lines	Agent reasoning traces

The simplest observability: structured logging

import json, time, os
from datetime import datetime

def log(event_type: str, **kwargs):
    entry = {"timestamp": datetime.utcnow().isoformat(), "event": event_type, **kwargs}
    print(json.dumps(entry))

def handler(event, context):
    run_id = os.urandom(4).hex()
    start = time.time()
    log("agent.start", run_id=run_id, goal=event.get("goal", ""))
    try:
        result = run_agent(event.get("goal", ""))
        log("agent.complete", run_id=run_id, duration_sec=round(time.time()-start, 2),
            status="success", result=result)
        return {"ok": True, "run_id": run_id}
    except Exception as e:
        log("agent.error", run_id=run_id, error=str(e))
        raise

Tool call tracing

def traced_tool_call(tool_name, args, run_id):
    start = time.time()
    log("tool.call", run_id=run_id, tool=tool_name, args=args)
    try:
        result = execute_tool(tool_name, args)
        log("tool.result", run_id=run_id, tool=tool_name,
            duration_ms=int((time.time()-start)*1000))
        return result
    except Exception as e:
        log("tool.error", run_id=run_id, tool=tool_name, error=str(e))
        raise

What the structured logs above reconstruct: one traced agent run as a waterfall. The LLM call dominates the runtime — without per-event timing you would never know.

Built-in observability with HollowHost

Metric	HollowHost dashboard	DIY logs
Run history	✅ Auto-tracked	You build it
Duration per run	✅ Dashboard	You compute it
Token usage	✅ Per-run tracking	You add code
Log streaming	✅ Real-time CLI	You pipe logs

hollowhost runs logs <ai-job-id> <run-id>

When to use a platform vs DIY

	DIY (this article)	Platform (LangSmith, Arize)
Cost	$0	$50-500+/month
Setup time	30 minutes	1-2 days
Custom metrics	Unlimited	Limited
Best for	1-5 agents, custom needs	10+ agents, teams

Start with DIY. Upgrade to a platform when you need shared dashboards.

---

Deploy agents with built-in observability. Every run is logged automatically. Start on HollowHost — free tier included.