End-to-end traces
Follow model calls, tools, errors, latency, and usage through the complete run that produced them.
The feedback loop for production agents.
A trace explains what happened. An evaluation tells you whether it met the expectation. Lens keeps both connected so production evidence can inform the next release.
Comparable evaluation
Keep suite, case, metric, release, dataset version, and trace identity together as one evaluation run.
Evidence-backed gates
Compare a candidate with its baseline and make regressions visible before they reach production.
Self-hosted control
Send native Anvia telemetry or supported OTLP traces to infrastructure and storage your team operates.
$docker compose up -dTurn one run into the next decision.
Start with what happened in production, diagnose the behavior, measure the candidate on stable cases, then let an explicit gate carry the evidence into release review.
- 01
Trace
Capture the agent, model, tool, release, latency, and usage involved in one production run.
- 02
Diagnose
Move from an alert or regression into the exact span, input boundary, and failure that changed.
- 03
Evaluate
Replay stable cases against a published dataset version and correlate every result to its trace.
- 04
Gate
Compare the candidate with its baseline, apply thresholds, and ship only with evidence.
The operating picture, in one place.
Explore an example overview with sample data. Change the time window and inspect token usage, latency, model efficiency, and tool health.
Production AgentsOverview
Overview
448 spans · 4 active users in this window
Sample data
Token usage
Input and output tokens over time
07 PM11 PM03 AM07 AM11 AM03 PM06 PM
- Input tokens
- Output tokens
Throughput and errors
Traces and LLM generations, with failed traces highlighted
07 PM11 PM03 AM07 AM11 AM03 PM06 PM
- Generations
- Traces
- Errors
Generation duration
P50 and P95 duration for generation observations
07 PM11 PM03 AM07 AM11 AM03 PM06 PM
- P50
- P95
Tokens by model
Share of total generation tokens
015K30K45K60K
Model efficiency
Usage, duration, and reliability by generation model
| Model | Generations | Token share | Input | Output | Total | Tokens / gen | P95 duration | Errors |
|---|---|---|---|---|---|---|---|---|
| gemini-2.5-flash | 48 | 25.5% | 29.3K | 23K | 52.3K | 1,089 | 1.76s | 2.1% |
| gpt-4.1 | 48 | 25.5% | 29.3K | 23K | 52.3K | 1,089 | 1.86s | 4.2% |
| gpt-4.1-mini | 48 | 25.4% | 29.2K | 22.9K | 52.1K | 1,085 | 1.58s | 4.2% |
| claude-sonnet-4 | 48 | 23.6% | 27.1K | 21.3K | 48.4K | 1,008 | 1.92s | 6.3% |
Services
Token load and trace health by service
| Service | Traces | Generations | Tokens | P95 duration | Errors |
|---|---|---|---|---|---|
| customer-support-agent | 28 | 86 | 94.3K | 3.12s | 9.4% |
| billing-copilot | 14 | 40 | 59.4K | 1.96s | 6.1% |
| research-agent | 11 | 33 | 30.8K | 4.80s | 14.2% |
| docs-qa | 11 | 33 | 20.5K | 1.44s | 13.0% |
Tool health
Most-used tool calls, duration, and failures
| Tool | Calls | P95 call duration | Errors |
|---|---|---|---|
| search_knowledge_base | 96 | 428ms | 2.1% |
| get_customer_context | 64 | 312ms | 0.0% |
| lookup_invoice | 32 | 186ms | 3.1% |
Token-heavy traces
Highest token usage in this window
Resolve a customer support requestgpt-4.1 · 2m ago7,230tokens
Summarize the research findingsgemini-2.5-flash · 5m ago6,418tokens
Explain an invoice adjustmentclaude-sonnet-4 · 8m ago5,804tokens
Recent failures
Latest traces with an error
Retrieve account permissionsgpt-4.1-mini · 3m ago1,240tokens
Search the support knowledge basegemini-2.5-flash · 11m ago2,186tokens
A score describes a case. A gate informs a release.
Lens groups the suite, immutable dataset version, metric direction, usage, outcomes, and trace references as one candidate that can be compared and gated.
Candidate
support-agent@1.4.0-candidate.2
Dataset support-regression@v12 · 48 cases · 192 metric results
- Passed
- 46
- Failed
- 2
- Eval usage
- 184k
MetricBaselineCandidateDelta
Answer relevancyHigher is better0.910.94+0.03pass
FaithfulnessHigher is better0.960.90−0.06fail
HallucinationLower is better0.040.07+0.03fail
Turn relevancyHigher is better0.880.90+0.02pass
Observe the agent. Flush the evidence.
@anvia/lens creates isolated telemetry providers, attaches to the runtime you already own, and exports correlated traces over OTLP HTTP with project-scoped credentials.
1import { Agent } from '@anvia/core'2import { LensClient } from '@anvia/lens'34const lens = new LensClient()5const tracing = lens.observer({ captureMode: 'safe' })67const agent = new Agent({8 id: 'support',9 model,10 observability: { observers: { tracing } },11})1213await agent.generate({ prompt: 'Summarize this ticket.' })14await lens.flush()15await lens.close()Useful by default. Explicit when sensitive.
Safe capture exports operational metadata without input or output bodies. Full payloads remain opt-in; redaction and size limits are configurable and governed by your application.
Your runtime
Defines the agent, release policy, credentials, and which environments or data classes may include payloads.
Isolated adapter
No global provider takeover. Payload capture is off until explicitly enabled.
project-scoped OTLP
Operational traces
Agent, model, tool, environment, release, latency, usage, and errors.
Evaluation evidence
Run lifecycle, cases, metrics, outcomes, usage, and direct trace references.
Storage you operate
Store project-scoped traces and evaluation evidence in your Lens deployment.
Use flush() for an explicit delivery checkpoint; call close() during final cleanup.
Start with the signal. End with the cause.
Three repeatable paths from an operational or quality signal to the run-level evidence needed to act.
Find a latency regression
Compare P95 by release and service, open the slow cohort, then inspect the model and tool spans consuming the budget.
- P95 alert
- release cohort
- slow trace
- span timing
Explain a tool-failure cluster
Group errors by service and tool, follow one representative run, and separate bad input from permissions, retries, or provider failure.
- error cluster
- tool span
- run context
- root cause
Resolve a failed quality gate
Open the candidate comparison, inspect the failing cases, jump to their traces, then rerun the same published dataset version.
- blocked gate
- failed case
- linked trace
- rerun
Debug locally. Learn in production.
Studio helps you understand an agent before it ships. Lens is an optional production companion that keeps the same runtime concepts visible after deployment. Your agent can also use another observability backend.
Local
Prompt, inspect, intervene
Exercise the real local agent, resolve approvals, inspect context, and replay the workflow.
Explore Studio ProductionObserve, evaluate, gate
Investigate deployed runs, compare stable evaluation suites, and carry evidence into release review.
Open the Lens guideSame agent and trace concepts, from your machine to production.
Run Lens on your own infrastructure.
Download the Compose and environment templates, configure unique secrets, then start the stack. The first account becomes the workspace owner.
- 01
Prepare the deployment
Download the official templates, pin a Lens release, and configure unique secrets.
- 02
Start the workspace
Launch the stack. The first account created becomes the workspace owner.
- 03
Connect a project
Create a project and ingestion key, then send native Anvia telemetry or supported Langfuse v5 OTLP traces.
Make the evidence part of the release.
Connect one agent in safe-capture mode, investigate its traces, then promote a stable evaluation suite into an explicit quality gate.
