AI agent latency & logging
Your agent takes 12 seconds to answer and nobody can say why. We trace every LLM call and tool call, cut the waiting time, and set up logging that tells you what your agent did, how long it took and what it cost.
before 9.4 s
after 2.1 s parallel tools · prompt cache · streaming
run.start user=[redacted] agent=supportllm.call model=fast in=1 204 out=88 cache=hittool.call name=search_orders 312 mstool.retry name=crm attempt=2 timeoutllm.call model=large in=3 980 out=412run.end 2.1 s · $0.0042 · 6 spansSound familiar?
01Ten seconds or more per answer. Tools run one after the other, prompts grow with every turn, and nothing streams.
02A customer complains about a wrong answer. You have no trace of the prompt, the tool results or the model's reasoning.
03LLM spend doubled this month. You can't tell which agent, which user or which loop is burning the tokens.
04A prompt tweak or a new model version ships on Friday. On Monday the agent is slower and less accurate, and nobody saw it happen.
Services
Start with an audit, get your tracing and logging right, then let us watch latency and cost month after month. Most teams combine all three.
Step in · 2 weeks
We trace your agent end to end on real traffic and find where the seconds go: sequential tool calls, oversized prompts, missing caching, the wrong model for the job. You get a ranked plan with the expected gain of each fix.
Build · 2–4 weeks
We instrument your agent so that every run leaves a full trace and clean, structured logs: prompts, tool calls, tokens, cost and errors, linked by one run ID, with personal data redacted before it leaves your servers.
Run · subscription
We watch your traces every week, check each prompt or model change against your evaluation set, and fix latency or quality regressions before your users notice. Each month you get a short report with the trends that matter.
Process
You show us the agent and what hurts. We tell you honestly if we can help and which offer fits.
NDA signed, a staging copy or a sampled slice of production traffic. Personal data is redacted at the source. We never change prompts in production without your approval.
Every finding is ranked by impact and effort: seconds saved, dollars saved, risk to answer quality. You see the gain before your team spends time on it.
We ship the fixes with your team, compare before and after on the same evaluation set, then stay on for as long as it is useful.
Stack
Already have a tracing tool? We build on it. Starting from zero? We set up an open, OpenTelemetry-based pipeline you can move anywhere.
Why LangSmith Data
Making an agent fast is not about a bigger GPU. It's about knowing exactly which span is slow and why. That needs good traces first, then experience with the fixes.
Every change comes with a before and after on your own traffic: p50 and p95 latency, cost per run, answer quality.
A faster model or a shorter prompt is only a win if the answers stay right. We check every fix against an evaluation set.
Personal data is redacted before it is logged. Traces can stay in your own cloud if you prefer.
Open standards, everything documented in your repository. Your team owns the dashboards. Stop whenever you want.
Plans
Fixed quote after the discovery call. No surprises on the invoice.
Audit
One-off · fixed fee
Tracing & Logging
Project · 2–4 weeks
Agent Ops
Monthly subscription
FAQ
Not in clear text. We set up redaction at the source, so names, emails, phone numbers and other personal data are masked before anything is logged. We can work on a staging copy or a small, sampled slice of production traffic.
LangGraph, LangChain, OpenAI Agents SDK, Claude Agent SDK, Vercel AI SDK, LlamaIndex, and hand-written agent loops in Python or TypeScript. If your code calls an LLM, we can trace it.
Not if it's done right. Many gains don't touch the answer at all: parallel tool calls, prompt caching, streaming. When a fix does change the model or the prompt, we compare quality on an evaluation set before it ships.
No. We build on what you have and fill the gaps: missing spans, logs that don't link to traces, no cost per run, no alerts. If you are starting from zero, we recommend an OpenTelemetry-based setup so you can change vendors later.
No. LangSmith Data is an independent service. We work with LangSmith and with the other tracing tools listed above, and recommend whichever fits your team.
We're based in Europe and work in English and French. We overlap with US East Coast business hours.
Contact
Send us a few lines about your agent and what hurts. We reply within one business day with times for a free 30-minute trace review.