NewAgent Observability — find failures before customers do.

Agent observability · Traces

Every session, in full

Every prompt, reply and tool call, kept in full. Find any conversation in under a second.

How to set it up

How traces work in Oodle

Send once. Then search, read, and follow the cause.

1 · Send

Send everything you already have.

Oodle takes what your stack already produces: OpenAI, Anthropic, LangChain, Pydantic AI, LiteLLM, the Langfuse SDK, or plain OpenTelemetry. No new SDK, and nothing is sampled on the way in.

agent observability · overview
envAll ×× ⌄modelAll ×× ⌄agent_nameAll ×× ⌄service_nameAll ×× ⌄
⌄ Overview

LLM Calls

0.00K

Total Cost

$0.000

Total Tokens

0K

Error Rate

0.00%

P95 Model Latency

0.0s

2 · Read

Read the whole conversation.

The system prompt, every reply, every tool call and what came back. Sentiment and cost on each turn, so you can see where a conversation went wrong, not just that it did.

agent observability · sessions
Session IDTracesSentiment
0bxF9Zq5-3gWH-4RNj-xHoo-Bkr4eKoEtQLZ1-0.82
CYG0O2dR-uNq6-4Lpp-sMpM-NoG0zZl0Pc6a1
5PgrnZYZ-Y6Ge-4U4r-1yAX-hq0QYUQcZMjt1+0.00
EMo8gVqV-j5ZB-4MQH-Dvfm-3ApMKVqzCswR1+0.46
7TDLf4K7-I4K4-4Bly-tHQp-0Xouoloy2bMs1-0.29
KIA4QEqL-Z1sg-4UPA-3SJ2-F75vI6U2DJ9u1+0.46
0bxF9Zq5… · order_support_agent · 6 turns-0.82
  • UserI have a problem with my order #4471, the camera arrived broken.
  • AssistantSorry to hear that. Let me look up the order.getOrderDetails · 340ms
  • Tool{ "status": "delivered", "eligible_for_return": true }
  • AssistantFiling the return now.initiateReturn · 503
  • AssistantYour return has been filed. You’ll get a label by email.

3 · Follow

Not every agent bug is in the agent.

The model call, the API request that started it and the database query the tool made are all in one trace. When a reply is slow or wrong, you follow that one trace to the cause.

one trace · api → agent → tool → database
POST /api/v1/chat6.4sOK
invoke_agent support-agent6.1sOK
chat gpt-4.10.9sOK
execute_tool lookup_order4.8sOK
SELECT orders postgres4.7sOK

seq scan on orders, 2.1M rows

chat gpt-4.10.4sOK

“Sorry for the wait, I found your order.”

The agent wasn't slow. The query under it was.

Try it on your agents

Two commands. No credit card. Every session kept from the first one.