AI observability,
illustrated.
Follow a request through an AI app. See how traces, evaluations, and alerts help you understand what happened, and whether the answer was good.
It starts with a question.
Someone asks your app, ‘Where’s my order?’ Behind that little message, an agent calls a model, looks up the order, and writes an answer.
Think of your app as a workshop. Each station has a job. A request moves between them until the work is done.
Connect your appFollow the work inside.
Foglamp records each AI call as a trace. Inside it, smaller steps called spans show the model calls and tools, with their inputs, outputs, and timing.
The order lookup says ‘delayed.’ The answer says ‘arrives today.’ Following the trace shows where those two stopped agreeing.
Explore tracesMeet the agent doing the work.
An agent uses instructions, models, and tools to work on a task. Give it a name, and Foglamp brings its runs, costs, and errors together.
Help the customer with their order. Keep your answer short.
Instructions change, too. Foglamp finds prompt versions in your recorded runs, so you can compare what changed and see which runs used each version.
Get to know your agentsSome jobs take a team.
One agent sorts the question. Another checks the order. A workflow run groups the AI calls that belong to one job, even when different agents do the work.
Your code controls the steps and branches. Add a workflow name and run ID, and Foglamp brings the recorded calls into one view.
See how workflows connectA conversation keeps going.
The customer asks a follow-up. A session connects the traces in the same conversation, so you can understand the experience across several messages.
Your order arrives today.
Are you sure? Tracking says it’s delayed.Workflows group calls for a job. Sessions group calls for a conversation. The same trace can belong to both.
See the whole conversationSmall calls add up.
Each model call takes time and uses tokens, the pieces of text a model reads and writes. Foglamp tracks usage, cost, and latency so you can see where they go.
Pull back to see patterns by model, agent, or customer. Then open an individual trace to understand an expensive or slow run.
Explore cost intelligenceAn answer can work and still be wrong.
Evaluations check the quality of an answer. Code checks test clear rules, like valid JSON. LLM judges assess things like helpfulness or whether an answer agrees with its sources.
The answer promises delivery today, but the order data says it is delayed.
The inspection bench checks copies of recorded traffic. Choose which traces or steps to sample, review their scores, and open the ones that failed.
Explore evaluationsKnow when something needs you.
Set a limit for cost, latency, error rate, or an eval’s pass rate. Foglamp checks recent traffic every minute and sends an email when a rule fires.
Alert below 85% over 1 hour
The pass rate is above your threshold. All quiet at the watchtower.
Start with the alert, inspect its context and relevant traces, then follow the work back to the problem. The rule returns to healthy when the metric recovers.
Explore alertsNow you can see the whole picture.
A little light on
your AI app.
Follow the work. Check the answers.
Know when something needs your attention.