AI Agent Observability

SENTINELTRACE

Know exactly why your agent gave that answer — every token, every tool call, every decision, in real time.

3 AgentsOTel SemconvConvex PersistenceOTLP Export
OTel Spec
Scroll

Live Span Feed

Every step.
Every token.
Every decision.

Each agent call emits an OpenTelemetry span with full GenAI semantic attributes — operation name, model, token counts, latency, and PII flags.

sentinel — span streamlive
Span NameStatusLatencyAgent
orchestrator.plan_query✓ ok312 msOrchestrator
researcher.fetch_documents✓ ok1842 msResearcher
researcher.tool_call:search✓ ok228 msResearcher
pii.scan_content✓ ok44 msSystem
judge.evaluate_response✓ ok903 msJudge
orchestrator.synthesize✓ ok571 msOrchestrator
memory.persist_results✓ ok88 msSystem
Orchestrator

Plans the query, coordinates the team, synthesizes the final answer.

Researcher

Runs a tool-use loop — search + fetch — until sufficient evidence is gathered.

Compliance Judge

Reviews the Researcher's output for accuracy and compliance before it is persisted.

Multi-Agent Team

Three agents.
One mission.

The Orchestrator delegates. The Researcher loops through tools until it has enough evidence. The Compliance Judge reviews everything before a single byte hits Convex.

Powered by Anthropic claude-sonnet-4-6

Capabilities

Enterprise-ready, out of the box.

PII Masking

Regex-based server-side scrub of emails, phones, SSNs, and card numbers before spans reach the wire.

Active Learning

Pattern detectors surface latency spikes, token bloat, error clusters, and tool-call loops across every run.

Memory Timeline

Convex persists every run. Diff badges highlight what changed since the last scan so nothing goes unnoticed.

OTLP Export

Ship spans to any OpenTelemetry-compatible collector — Jaeger, Grafana Tempo, Honeycomb — with one env var.

How It Works

Three steps from query to insight.

01

Submit a Query

Type any prompt into the Query Composer. The Orchestrator receives it and plans which tools and agents to activate.

02

Agents Run & Emit Spans

Each agent step emits a real OpenTelemetry span tagged with gen_ai.system, model, token counts, latency, and PII flags. Watch the timeline update live.

03

Persist & Learn

The Compliance Judge validates the run, then everything is saved to Convex. Pattern detectors surface latency drift, error clusters, and PII recurrence across all historical runs.

Ready to trace
your agents?

Open the live dashboard — no setup required, mock stream starts instantly.