agent-observability – skill Agent & Tự động hóa cho agent AI

Agent Observability

Monitor AI agent behavior with logs, traces, metrics, and cost telemetry. This skill covers the full observability stack for LLM-powered applications: from raw Prometheus counters to Grafana dashboards, OpenTelemetry tracing, structured logging, cost tracking, SLO definition, and PII redaction.

---

Examples use custom llm.* attributes, not a claim of current OpenTelemetry semantic-convention compliance. Import shared metric definitions once and use bounded configured label values. Telemetry export is data disclosure: approve destination/retention and omit prompts, credentials, direct user IDs and raw exception bodies. TLS is required for non-loopback collectors.

Core Metrics

Define these metrics at the application layer. All examples use the Prometheus client library naming conventions.

Latency

from prometheus_client import Histogram

# Total end-to-end latency for a full agent turn (user prompt -> final response)
AGENT_LATENCY = Histogram(
    "agent_request_duration_seconds",
    "End-to-end latency of an agent request",
    labelnames=["agent_name", "model", "status"],
    buckets=(0.25, 0.5, 1, 2, 5, 10, 30, 60, 120),
)

# Latency of a single LLM API call (one completion request)
LLM_CALL_LATENCY = Histogram(
    "llm_call_duration_seconds",
    "Latency of an individual LLM API call",
    labelnames=["model", "provider", "stream"],
    buckets=(0.1, 0.25, 0.5, 1, 2, 5, 10, 30),
)

# Latency of tool/function calls executed by the agent
TOOL_CALL_LATENCY = Histogram(
    "agent_tool_call_duration_seconds",
    "Latency of a tool call executed by the agent",
    labelnames=["tool_name", "agent_name", "status"],
    buckets=(0.05, 0.1, 0.25, 0.5, 1, 2, 5, 10),
)

Token Usage

from prometheus_client import Counter, Histogram

PROMPT_TOKENS = Counter(
    "llm_prompt_tokens_total",
    "Total prompt tokens sent to the model",
    labelnames=["model", "agent_name"],
)

COMPLETION_TOKENS = Counter(
    "llm_completion_tokens_total",
    "Total completion tokens received from the model",
    labelnames=["model", "agent_name"],
)

CACHED_TOKENS = Counter(
    "llm_cached_tokens_total",
    "Prompt tokens served from KV-cache (provider-reported)",
    labelnames=["model", "agent_name"],
)

TOKENS_PER_REQUEST = Histogram(
    "llm_tokens_per_request",
    "Total tokens (prompt + completion) per request",
    labelnames=["model", "agent_name"],
    buckets=(100, 500, 1000, 2000, 4000, 8000, 16000, 32000, 64000, 128000),
)

Cost

from prometheus_client import Counter

LLM_COST = Counter(
    "llm_cost_dollars_total",
    "Estimated cost in USD for LLM usage",
    labelnames=["model", "agent_name", "cost_type"],  # cost_type: prompt | completion
)

Tool Calls

from prometheus_client import Counter

TOOL_CALLS_TOTAL = Counter(
    "agent_tool_calls_total",
    "Total tool calls made by agents",
    labelnames=["tool_name", "agent_name", "status"],  # status: success | error | timeout
)

Errors and Retries

from prometheus_client import Counter, Gauge

LLM_ERRORS = Counter(
    "llm_errors_total",
    "Errors returned by the LLM provider",
    labelnames=["model", "provider", "error_type"],  # error_type: rate_limit | timeout | 5xx | auth
)

LLM_RETRIES = Counter(
    "llm_retries_total",
    "Retried LLM API calls",
    labelnames=["model", "provider", "retry_reason"],
)

AGENT_ACTIVE_REQUESTS = Gauge(
    "agent_active_requests",
    "Number of agent requests currently in flight",
    labelnames=["agent_name"],
)

---

OpenTelemetry Integration

Use the OpenTelemetry Python SDK to create traces that capture every step of an agent turn: the top-level request, each LLM call, each tool execution, and retrieval operations.

Setup

# otel_setup.py
from opentelemetry import trace
from opentelemetry.sdk.trace import TracerProvider
from opentelemetry.sdk.trace.export import BatchSpanProcessor
from opentelemetry.exporter.otlp.proto.grpc.trace_exporter import OTLPSpanExporter
from opentelemetry.sdk.resources import Resource

def init_tracing(service_name: str, otlp_endpoint: str = "http://localhost:4317"):
    resource = Resource.create({
        "service.name": service_name,
        "service.version": "1.0.0",
        "deployment.environment": "production",
    })
    provider = TracerProvider(resource=resource)
    exporter = OTLPSpanExporter(endpoint=otlp_endpoint, insecure=otlp_endpoint.startswith("http://localhost:"))
    provider.add_span_processor(BatchSpanProcessor(exporter))
    trace.set_tracer_provider(provider)
    return trace.get_tracer(service_name)

Tracing LLM Calls

# llm_tracing.py
import time
from opentelemetry import trace
from opentelemetry.trace import StatusCode

tracer = trace.get_tracer("agent.llm")

def traced_llm_call(client, messages, model="gpt-4o", **kwargs):
    """Wrap an LLM completion call with a full OpenTelemetry span."""
    if kwargs.get("stream"):
        raise ValueError("This synchronous wrapper does not consume streams; use a stream-aware adapter")
    with tracer.start_as_current_span("llm.chat_completion", record_exception=False, set_status_on_exception=False) as span:
        span.set_attribute("llm.model", model)
        span.set_attribute("llm.provider", "openai")
        span.set_attribute("llm.message_count", len(messages))
        span.set_attribute("llm.temperature", kwargs.get("temperature", 1.0))
        span.set_attribute("llm.max_tokens", kwargs.get("max_tokens", 0))

        start = time.perf_counter()
        try:
            response = client.chat.completions.create(
                model=model, messages=messages, **kwargs
            )
            elapsed = time.perf_counter() - start

            usage = response.usage
            if usage is None or not response.choices:
                raise ValueError("Usage/choices unavailable; record unknown rather than inventing zero")
            span.set_attribute("llm.prompt_tokens", usage.prompt_tokens)
            span.set_attribute("llm.completion_tokens", usage.completion_tokens)
            span.set_attribute("llm.total_tokens", usage.total_tokens)
            span.set_attribute("llm.duration_seconds", elapsed)
            span.set_attribute("llm.finish_reason", response.choices[0].finish_reason)
            span.set_status(StatusCode.OK)

            # Update Prometheus counters
            PROMPT_TOKENS.labels(model=model, agent_name="default").inc(usage.prompt_tokens)
            COMPLETION_TOKENS.labels(model=model, agent_name="default").inc(usage.completion_tokens)
            LLM_CALL_LATENCY.labels(model=model, provider="openai", stream="false").observe(elapsed)

            return response

        except Exception as exc:
            elapsed = time.perf_counter() - start
            span.set_status(StatusCode.ERROR, type(exc).__name__)
            # Raw exception text may contain prompts, URLs or credentials.
            LLM_ERRORS.labels(model=model, provider="openai", error_type="request_error").inc()
            LLM_CALL_LATENCY.labels(model=model, provider="openai", stream="false").observe(elapsed)
            raise

Tracing Tool Execution

# tool_tracing.py
import functools
from opentelemetry import trace
from opentelemetry.trace import StatusCode

tracer = trace.get_tracer("agent.tools")

def traced_tool(tool_name: str):
    """Decorator that wraps a tool function with an OTel span and Prometheus metrics."""
    def decorator(func):
        import inspect
        if inspect.iscoroutinefunction(func) or inspect.isgeneratorfunction(func):
            raise TypeError("Use an async/generator-aware adapter that observes actual completion")
        @functools.wraps(func)
        def wrapper(*args, **kwargs):
            with tracer.start_as_current_span(f"tool.{tool_name}", record_exception=False, set_status_on_exception=False) as span:
                span.set_attribute("tool.name", tool_name)
                span.set_attribute("tool.args_count", len(args) + len(kwargs))

                import time
                start = time.perf_counter()
                try:
                    result = func(*args, **kwargs)
                    elapsed = time.perf_counter() - start
                    span.set_attribute("tool.duration_seconds", elapsed)
                    span.set_status(StatusCode.OK)
                    TOOL_CALLS_TOTAL.labels(
                        tool_name=tool_name, agent_name="default", status="success"
                    ).inc()
                    TOOL_CALL_LATENCY.labels(
                        tool_name=tool_name, agent_name="default", status="success"
                    ).observe(elapsed)
                    return result
                except Exception as exc:
                    elapsed = time.perf_counter() - start
                    span.set_status(StatusCode.ERROR, type(exc).__name__)
                    # Raw exception text may c

… (nội dung đầy đủ có trong gói tải về)