Batch logging for your agent

For frameworks that have already completed the LLM call and only need to record it, use tracing.log_turn and tracing.log_conversation from the CoreWeave Forge SDK. All spans are created and ended immediately without keeping any context managers open. Data logged this way can be historical. No live conversation is needed. Set conversation_id to any stable string that uniquely identifies the conversation. Turns that share the same conversation_id are grouped as a single conversation in the Conversations tab. If you’re building your own agent loop, use the real-time instrumentation APIs described in Trace your agents instead.

Log a turn

To record a single completed turn after it happens, use tracing.log_turn. It accepts a fully-formed turn, including all LLM and tool spans. Construct the LLM and Tool objects directly rather than through a start_* factory. A directly constructed span doesn’t emit anything until you pass it to log_turn, so these objects work as plain data.

Python
TypeScript

from coreweave.forge.agentlens import tracing
from coreweave.forge.agentlens.tracing import LLM, Message, Tool, Usage

tracing.init("[YOUR-TEAM]/[YOUR-PROJECT]")

llm_span = LLM(
    model="gpt-4o",
    provider_name="openai",
    input_messages=[Message(role="user", content="What is the weather?")],
    output_messages=[Message(role="assistant", content="Let me check.")],
    usage=Usage(input_tokens=100, output_tokens=20),
)

tool_span = Tool(
    name="get_weather",
    arguments='{"city": "Tokyo"}',
    result='"24°C, sunny"',
)

llm_span2 = LLM(
    model="gpt-4o",
    provider_name="openai",
    input_messages=[Message(role="user", content="What is the weather?")],
    output_messages=[Message(role="assistant", content="It is 24°C and sunny.")],
    usage=Usage(input_tokens=150, output_tokens=30),
)

# Log a turn with all its spans.
tracing.log_turn(
    conversation_id="my-conversation-abc",
    agent_name="weather-bot",
    messages=[
        Message(role="user", content="What is the weather in Tokyo?"),
        Message(role="assistant", content="It is 24°C and sunny in Tokyo."),
    ],
    spans=[llm_span, tool_span, llm_span2],
)

This feature is not available in the TypeScript SDK yet.

log_turn returns a LogResult containing the trace IDs of the emitted spans. An optional model parameter on log_turn sets the model on the turn’s own span, not on the child LLM spans. Each LLM span carries its own model independently. If a turn uses multiple models, set model on log_turn to whichever you consider the primary model for that turn. To record when the turn actually happened rather than when you logged it, pass started_at and ended_at as timezone-aware datetime values.

Log a conversation

To bulk-import a complete, multi-turn conversation at once, use tracing.log_conversation. The turns parameter accepts a list of Turn objects, each constructed the same way as the spans in the previous log_turn example.

Python
TypeScript

tracing.log_conversation(
    conversation_id="my-conversation-abc",
    agent_name="weather-bot",
    turns=[turn_1, turn_2],
)

This feature is not available in the TypeScript SDK yet.

​Log a turn

​Log a conversation

Log a turn

Log a conversation