agenx

Observability and tracing for AI agents, LLM chains, and workflows. Trace every step of your AI pipeline — LLM calls, tool use, retrieval, guardrails — with zero boilerplate.

Installation

pip install agenx

With LangChain support:

pip install agenx[langchain]

All optional dependencies:

pip install agenx[all]

Requires Python 3.11+.

Quick start

Decorators

The fastest way to add tracing. Decorate your functions and agenx handles span creation, timing, and parent-child relationships automatically.

from agenx import Tracer, StdoutExporter

tracer = Tracer()
tracer.configure(exporters=[StdoutExporter()])

@tracer.llm(model="gpt-4", provider="openai", temperature=0.7)
def chat(prompt: str) -> str:
    return call_openai(prompt)

@tracer.tool(tool_name="web_search")
def search(query: str) -> list:
    return fetch_results(query)

@tracer.agent(tools=["search", "chat"])
def assistant(question: str) -> str:
    docs = search(question)
    return chat(f"Answer using: {docs}")

assistant("What is quantum computing?")

Context managers

For manual control over span attributes and events:

from agenx import trace, SpanType

with trace.span("pipeline") as root:
    root.step.input = {"user_id": 12}

    with trace.span("search", type=SpanType.RETRIEVER) as r:
        r.retriever.query = "reset password"
        r.retriever.returned = 5
        r.add_event("cache_hit", {"key": "search_abc"})

    with trace.span("generate", type=SpanType.LLM) as llm:
        llm.llm.model = "gpt-4o"
        llm.llm.tokens_total = 843

    root.step.output = {"answer": "Click forgot password on the login page"}

trace is a pre-configured Tracer instance with StdoutExporter enabled.

Span types

Each decorator creates a span with type-specific attributes:

Decorator options

All decorators accept:

• name — Custom span name (defaults to function name)
• capture_args — Record function arguments (default: True)
• capture_result — Record return value (default: True)

# Disable capture for sensitive operations
@tracer.step(capture_args=False, capture_result=False)
def handle_credentials(api_key: str) -> dict:
    ...

Type-specific options:

@tracer.llm(model="claude-sonnet-4-6", provider="anthropic", temperature=0.3)
@tracer.agent(tools=["search", "calculator"])
@tracer.tool(tool_name="math_calculator")
@tracer.retriever(top_k=10)
@tracer.guard(guard_name="content_safety_filter")

Async support

All decorators work with both sync and async functions:

@tracer.llm(model="claude-sonnet-4-6", provider="anthropic")
async def generate(prompt: str) -> str:
    return await async_llm_call(prompt)

Async context managers are also supported:

async with tracer.aspan("async_pipeline") as span:
    span.step.input = {"query": "hello"}

Exporters

StdoutExporter

Built-in colored terminal output with hierarchy and timing:

from agenx import StdoutExporter

exporter = StdoutExporter(
    indent=True,           # Show parent-child hierarchy
    show_ids=True,         # Display trace/span IDs
    min_duration_ms=0.0,   # Filter by minimum duration
)

Custom exporters

Any callable that accepts a Span works as an exporter:

def my_exporter(span):
    print(f"{span.name} took {span.to_dict().get('duration_ms', 0):.1f}ms")

tracer = Tracer()
tracer.configure(exporters=[StdoutExporter(), my_exporter])

Automatic nesting

Spans nest automatically via context. No manual ID passing needed:

@tracer.agent(name="qa_agent")
def qa_pipeline(question: str) -> str:
    docs = retrieve(question)       # child span
    answer = generate(question, docs)  # child span
    return answer

@tracer.retriever(top_k=5)
def retrieve(query: str) -> list:
    ...

@tracer.llm(model="gpt-4")
def generate(query: str, context: list) -> str:
    ...

Calling qa_pipeline produces a trace tree:

qa_agent
  ├── retrieve
  └── generate

LangChain integration

Auto-instrument LangChain and LangGraph with a single call:

from agenx import instrument

instrument("langchain")

This patches chains, LLM calls, agents, and tools. To instrument all detected frameworks:

from agenx import instrument_all

instrument_all()

To remove instrumentation:

from agenx import uninstrument, uninstrument_all

uninstrument("langchain")
# or
uninstrument_all()

Events

Record point-in-time events within a span:

with trace.span("search", type=SpanType.RETRIEVER) as r:
    r.add_event("cache_hit", {"key": "search_abc"})
    r.add_event("reranking_complete", {"top_k": 5})

API reference

Exports from agenx

License

MIT