AI Tracing for
Production Agents
Capture agent and LLM spans with the Respan SDK or OTLP. Debug waterfalls, roll up cost by feature, and score evals on the trace tree.
See exactly what your agents did.
Every LLM call, tool run, retrieval, and agent turn becomes a span in one trace, with its input, output, latency, and cost captured.
Every step of a request, in one trace
LLM calls, tool runs, retrievals, and agent turns each become a span in one trace, showing exactly where a run's time and cost go.
{
"model": "claude-opus-4-8",
"temperature": 0.2,
"max_tokens": 1024,
"messages": [
{ "role": "system", "content": "You are a support agent. Answer only from the retrieved context." },
{ "role": "user", "content": "Why was I charged twice this month?" },
{ "role": "assistant", "content": "Let me check your recent invoices and payment retries." }
]
}Any span, down to the last field
Model, latency, cost, tokens, and the exact input and output sit on every span, showing precisely what each step ran and returned.
Score production, not just tests.
Evaluators run on sampled live traffic and their scores land on the spans you debug, so a drop reaches you as an alert, not a ticket.
Scores on the trace itself
Every evaluator result attaches to the span that produced it, so you can filter to the runs that scored badly and open them.
The same evaluators on live traffic
Run your evaluators against sampled production traffic, so quality is measured on what users actually sent, not a fixed test set.
An alert when scores fall
Set a threshold on any evaluator and Respan tells you when live scores cross it, instead of waiting for a customer to report it.
Logs cannot replay a run.
Flat records lose the shape of a request, so the step that failed, the tool that stalled, and what it cost all go missing.
Every call is its own record
Each model and tool call lands separately, so a ten step run arrives as ten unrelated entries with nothing to say which came from which.
The tool runs never show up
Only the model calls get recorded, so retrievals and function steps are missing, and the slow step is never the one you can actually see.
The trail stops at the service edge
Retrieval runs in one service and orchestration in another, so a request that crosses between them arrives as two histories that never meet.
Sampled away before anyone looked
Traffic gets thinned at the door to hold down storage, and the rare broken run nobody has seen yet is exactly the one that gets dropped.
Customer data kept without a plan
Prompts and replies are stored exactly as they were sent, so names, emails, and account numbers pile up because nobody decided otherwise.
Quality scores sit somewhere else
Results live in a spreadsheet or a separate tool, so a run that scored badly cannot be opened and read beside what actually produced it.
Start tracing in a few lines.
Call Respan() once and the SDKs you already use are traced, then add a decorator when you want a multi-step run to arrive as one tree.
from openai import OpenAIfrom respan import Respanfrom respan.decorators import workflowfrom respan.contexts.span import respan_span_attributes Respan()client = OpenAI() @workflow(name="support_agent")def support_agent(question): with respan_span_attributes( respan_params={"customer_identifier": "user_123"} ): return client.chat.completions.create( model="claude-opus-4-8", messages=[{"role": "user", "content": question}], )from openai import OpenAIfrom respan import Respanfrom respan.decorators import workflowfrom respan.contexts.span import respan_span_attributes Respan()client = OpenAI() @workflow(name="support_agent")def support_agent(question): with respan_span_attributes( respan_params={"customer_identifier": "user_123"} ): return client.chat.completions.create( model="claude-opus-4-8", messages=[{"role": "user", "content": question}], )The AI observability platform behind 80 trillion+ tokens.
Loved by world-class founders, engineers, and product teams.
“Imagine jumping to a log immediately after every LLM call. This is the dream for debugging.”
Daniel Wolf
Product Lead, AlphaSense
“We scaled from 5M to 500M+ monthly API calls quickly. Respan gave us the debugging layer to resolve production issues 10x faster.”
Read how Retell builds next-gen voice agents that scale->Zexia Zhang
CTO, Retell AI
“Respan legit has some of the best UX/DX I’ve ever seen in my life. I truly don’t think I’ve ever integrated a product that was as easy.”
Rahul Behal
Co-founder, Gumloop
“This one felt pretty nice.”
Fabian Hedin
CTO, Lovable
“Such a no brainer choice over LangSmith or anything else and super easy to set up.”
Andy Wang
CEO, Finta
“Respan has been key in helping us scale to trillions of tokens reliably with real-time observability.”
Read how Mem0 builds reliable self-improving AI memory layer->Deshraj Yadav
CTO, Mem0
“Great product - really love the metrics dashboard.”
Esha Dinne
CTO, Giga
Certified where it counts.
Built to meet the security and privacy standards that enterprise and healthcare teams require.
ISO 27001
The internationally recognized standard for information security management.
SOC 2
Secure, compliant management of your data across all of our systems.
GDPR
Operated under GDPR, the world's strictest standard for data privacy.
HIPAA
HIPAA compliant, with a BAA available for healthcare teams.
Works with your entire stack
Use Respan with your favorite frameworks and tools.
