Observability

Trace Observatory · every agent call, every token

OpenTelemetry GenAI instrumentation flows into Langfuse SDK (server-side capture) and renders here using Vercel AI Elements + assistant-ui primitives. Token/cost/latency per call · drill in for full reasoning chain.

📡Traces · 24h

284

across all workflows

⏱️p50 latency

4.2s

p95 12.5s

🪙Tokens · 24h

2.1M

1.5M in · 600k out

💰Cost · 24h

$5.10

claude-opus-4-7 + sonnet-4-6

Latency

p50 / p95 over the last hour

Cost

Per-agent · last 24h ($USD)

Recent traces

Click to drill in

Trace detail

trace-9847

Workflow

Reconciliation

Agent

coordinator

Started

14:22:18

Duration

12.4s

Tool calls

Tokens in

8,420

Tokens out

1,180

Cost

$0.062

Reasoning chain (sample)

▸ coordinator · plan "Custody + Accounting in parallel, then Recon for variance attribution" ▸ tool_use · fireblocks__get_vault_balance args: { vault_id: "KFCustody-USDC-Hot-01", asset: "USDC" } result: 50,000,000.00 ▸ tool_use · netsuite__get_gl_balance args: { account: "10520" } result: 48,750,000.00 ▸ coordinator · reasoning "Variance $1.25M between custody and GL. Checking on-chain." ▸ tool_use · etherscan__get_gas_used args: { tx_hash: "0xae3f8b…c4d2" } result: $1,247,890.30 cumulative ▸ coordinator · conclusion "Variance = gas_leakage · confidence 0.94 · post adj JE"

Links

📓 Audit row 0x8f1c…d3a2 🔗 Open in Langfuse

Stack

What this view is built on

Capture: @vercel/otel + Langfuse SDK on every Claude Agent SDK call · Render: Vercel AI Elements (reasoning blocks · tool-call cards) + assistant-ui (Thread, Message, ToolCall primitives) + Recharts (token/cost/latency) + TanStack Table (trace list) · Persistence: Langfuse (self-hosted at langfuse.vaultprotocol.ai) for off-app deep-link.