agensights 0.3.0

AgenSights SDK - AI Agent Observability. Zero-friction tracking for LLM calls, agents, and tools.

AgenSights Python SDK

Python SDK for AgenSights - AI Agent Observability.

Track LLM calls, tool invocations, and multi-step agent executions with zero-friction auto-instrumentation or manual tracking.

Installation

pip install agensights

Or install from source:

pip install -e .

Quick Start — Universal Init (Recommended)

One line at the top of your app patches every supported LLM provider automatically:

import agensights

agensights.init(api_key="sk-dev-xxx")

# That's it. Every OpenAI, Anthropic, Bedrock, Google, Mistral,
# Cohere, and LiteLLM call is now tracked automatically.

You can also configure via environment variables (no code changes needed):

export AGENSIGHTS_API_KEY="sk-dev-xxx"
export AGENSIGHTS_BASE_URL="https://api.agensights.dev/api/v1"

import agensights
agensights.init()  # picks up from env vars

Auto-Instrumentation (Per-Client)

Wrap your LLM client once and every call is tracked automatically.

OpenAI

from openai import OpenAI
from agensights import instrument_openai

client = instrument_openai(
    OpenAI(api_key="sk-xxx"),
    agensights_api_key="sk-dev-xxx",
    agent_name="my-assistant",
)

# Every call is now automatically tracked
response = client.chat.completions.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": "Hello!"}],
)

# Embeddings are tracked too
embeddings = client.embeddings.create(
    model="text-embedding-3-small",
    input="Hello world",
)

Anthropic

import anthropic
from agensights import instrument_anthropic

client = instrument_anthropic(
    anthropic.Anthropic(api_key="sk-ant-xxx"),
    agensights_api_key="sk-dev-xxx",
    agent_name="claude-agent",
)

# Automatically tracked
message = client.messages.create(
    model="claude-3-5-sonnet-20241022",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello!"}],
)

LangChain

from langchain_openai import ChatOpenAI
from agensights.integrations import LangChainCallbackHandler

handler = LangChainCallbackHandler(
    api_key="sk-dev-xxx",
    agent_name="langchain-agent",
)

llm = ChatOpenAI(model="gpt-4o", callbacks=[handler])

# All LLM and tool calls are tracked via callbacks
response = llm.invoke("Hello!")

Agent Hierarchy Tracking

Track multi-agent workflows with automatic parent-child relationships:

from agensights import instrument_openai
from openai import OpenAI

client = instrument_openai(OpenAI(), agensights_api_key="sk-dev-xxx")

with client.trace("find_laptop") as trace:
    with trace.agent("planner") as planner:
        with planner.agent("researcher") as researcher:
            with researcher.tool("web_search"):
                results = do_search("laptops")  # latency auto-measured
            # LLM call auto-captured under researcher agent
            summary = client.chat.completions.create(
                model="gpt-4o",
                messages=[{"role": "user", "content": f"Summarize: {results}"}],
            )
        with planner.agent("writer") as writer:
            result = client.chat.completions.create(
                model="gpt-4o",
                messages=[{"role": "user", "content": "Write recommendation"}],
            )

This produces a full trace tree in the dashboard with parent-child spans linked automatically.

Manual Tracking

For full control, use the AgenSights client directly.

Single Calls

from agensights import AgenSights

client = AgenSights(api_key="sk-prod-xxx")

# Track a single LLM call
client.track_llm(model="gpt-4o", input_tokens=100, output_tokens=50, latency_ms=300)

# Track a tool call
client.track_tool(tool_name="web_search", latency_ms=150)

# Always close when done
client.close()

Tracing Multi-Step Executions

Use client.trace() to group related calls under a single trace:

from agensights import AgenSights

client = AgenSights(api_key="sk-prod-xxx")

with client.trace("support_agent", workflow_id="ticket-456") as t:
    # Track an LLM call
    t.llm_call(model="gpt-4o", input_tokens=100, output_tokens=50, latency_ms=300)

    # Track a tool call
    t.tool_call(tool_name="web_search", latency_ms=150)

    # Use spans for automatic duration tracking
    with t.span("data_processing") as s:
        # ... your code here ...
        pass  # duration is recorded automatically

client.close()

Nested Agent Spans

with client.trace("orchestrator") as t:
    planner = t.agent("planner")
    researcher = planner.agent("researcher")  # sub-agent
    researcher.tool(name="search_api", latency_ms=150)
    researcher.llm_call(model="gpt-4o", input_tokens=100, output_tokens=50, latency_ms=300)

    writer = planner.agent("writer")
    writer.llm_call(model="claude-3-5-sonnet", input_tokens=200, output_tokens=100, latency_ms=400)

Using the Client as a Context Manager

with AgenSights(api_key="sk-prod-xxx") as client:
    client.track_llm(model="gpt-4o", input_tokens=100, output_tokens=50, latency_ms=300)
# Client is automatically closed and flushed

Configuration

Environment Variables

Variable	Description
AGENSIGHTS_API_KEY	Your AgenSights API key (used when api_key is not passed)
AGENSIGHTS_BASE_URL	Backend API base URL (default: https://api.agensights.com/api/v1)

Client Parameters

Parameter	Default	Description
api_key	AGENSIGHTS_API_KEY env var	Your AgenSights API key
base_url	AGENSIGHTS_BASE_URL env var	Backend API base URL

Auto-Instrumentation Parameters

Parameter	Default	Description
agensights_api_key	None	API key (or pass agensights_client instead)
agensights_client	None	Pre-configured AgenSights instance
agent_name	None	Name to tag all events with
base_url	None	Override backend URL (falls back to env var)

Error Tracking

Errors are automatically captured during auto-instrumentation. For manual tracking:

client.track_llm(
    model="gpt-4o",
    input_tokens=100,
    output_tokens=0,
    latency_ms=500,
    status="error",
    error_code="rate_limit",
)

How It Works

• Universal init (agensights.init()) patches all supported LLM providers at the module level.
• Auto-instrumentation wraps LLM client methods (e.g., chat.completions.create) to capture model, tokens, latency, and errors transparently.
• Events are buffered locally and sent in batches to the AgenSights backend.
• The buffer flushes automatically every 5 seconds or when 100 events are accumulated.
• Call client.flush() to force an immediate send.
• Call client.close() to flush and release resources.

Supported Providers

Provider	agensights.init()	instrument_*()
OpenAI	Auto-patched	instrument_openai()
Anthropic	Auto-patched	instrument_anthropic()
AWS Bedrock	Auto-patched	via init()
Google Gemini	Auto-patched	via init()
Mistral AI	Auto-patched	via init()
Cohere	Auto-patched	via init()
LiteLLM	Auto-patched	via init()
LangChain	—	LangChainCallbackHandler
CrewAI	—	CrewAITracker
AutoGen	—	AutoGenTracker
Google ADK	—	GoogleADKTracker

Development

pip install -e ".[dev]"
pytest

License

MIT - see LICENSE for details.