Dashboard Overview

Dashboard Overview

The Opswald dashboard is your central hub for debugging AI agents. View traces, replay sessions, and understand why your agents behave the way they do.

Getting Started

Access your dashboard at app.opswald.com after setting up traces with the SDK or Proxy.

First Session

When you first instrument your AI agent, you’ll see:

1. Real-time traces appear as your agent runs
2. Session timeline showing each step in chronological order
3. Quick insights highlighting errors, slow steps, and expensive calls

Main Navigation

Traces

View all your agent execution traces:

• Live traces - Real-time view of running agents
• Recent sessions - Last 50 completed agent runs
• Search & filter - Find specific traces by date, model, cost, duration

Replay Lab

Debug by replaying agent sessions:

• Interactive replay - Step through any trace in slow motion
• Fork & modify - Change inputs and see different outcomes
• Golden tests - Pin important sessions as regression benchmarks

Decision Graphs

Visualize agent reasoning:

• Decision tree - See the full graph of choices your agent considered
• Critical path - Highlight the sequence that led to the final outcome
• Ghost paths - Visualize alternatives the agent didn’t take

Session Timeline

Every agent session is displayed as an interactive timeline:

08:15:23  🤖 Agent Started    session_id: chat-user-456
08:15:24  📝 User Message     "Analyze this sales report"
08:15:25  🔍 Tool Call       read_file(sales-q1.csv)
08:15:26  📊 LLM Call        gpt-4o (1.2k tokens, $0.018)
08:15:28  🔍 Tool Call       create_chart(data, type='bar')
08:15:30  ✅ Agent Complete  Total: $0.018, 3.2s

Click any step to see:

• Full request/response data
• Token counts and costs
• Latency breakdown
• Error details (if any)

Real-Time Monitoring

Watch your agents as they run:

Live Trace View

• Active sessions appear immediately
• Step-by-step progress updates in real-time
• Performance metrics (latency, cost) update live
• Error highlighting when things go wrong

Status Indicators

• 🟢 Running - Agent actively processing
• 🔵 Waiting - Agent waiting for user input or tool response
• 🟡 Slow - Step taking longer than usual
• 🔴 Error - Step failed or agent crashed

Debugging Features

Error Analysis

When your agent fails, the dashboard shows:

🔴 Error in Step 3: Tool Call Failed
Tool: send_email
Error: SMTP authentication failed
Context: Sending weekly report to team@company.com
Suggested fix: Check SMTP credentials in environment

Performance Insights

Identify bottlenecks:

• Slowest steps highlighted in red
• Most expensive calls marked with 💰
• Token efficiency metrics per model
• Latency distribution across tool calls

Cost Breakdown

Track spending per session:

Session Cost Breakdown
├── LLM Calls: $0.145 (89%)
│   ├── gpt-4o: $0.098 (3 calls)
│   └── claude-3-sonnet: $0.047 (1 call)
├── Tool Calls: $0.018 (11%)
│   ├── Web search: $0.012
│   └── File storage: $0.006
└── Total: $0.163

Search & Filtering

Find specific traces quickly:

Quick Filters

• Today - All traces from today
• Errors - Only failed sessions
• Expensive - Sessions costing >$1
• Long running - Sessions >30 seconds

Advanced Search

model:gpt-4o cost:>0.50 duration:<10s user:john error:false

Supported filters:

• model: - Filter by LLM model
• cost: - Filter by session cost ($)
• duration: - Filter by total time
• user: - Filter by user ID
• error: - true/false for error presence
• tool: - Filter by tool calls made

Session Management

Bulk Operations

• Delete sessions - Remove test or sensitive data
• Export traces - Download for external analysis
• Tag sessions - Organize with custom labels

Privacy Controls

• Content filtering - Hide sensitive data while preserving structure
• PII redaction - Automatically mask personal information
• Retention settings - Auto-delete old traces

Keyboard Shortcuts

Navigate faster with shortcuts:

Shortcut	Action
?	Show all shortcuts
/	Focus search
j/k	Next/previous trace
Enter	Open selected trace
Esc	Close modals
r	Refresh trace list

Mobile Experience

The dashboard is optimized for mobile debugging:

• Responsive design works on phones and tablets
• Touch gestures for timeline scrubbing
• Condensed view shows essential information
• Quick actions accessible via swipe

Integration

Slack Alerts

Get notified when agents fail:

# Set up Slack webhooks
curl -X POST https://api.opswald.com/v1/integrations/slack \
  -H "Authorization: Bearer your-api-key" \
  -d '{
    "webhook_url": "https://hooks.slack.com/...",
    "alerts": ["error", "high_cost", "long_duration"]
  }'

API Access

Export data programmatically:

from opswald import OpsClient

client = OpsClient("your-api-key")

# Get recent traces
traces = client.traces.list(limit=50, error=False)

for trace in traces:
    print(f"Session: {trace.session_id}")
    print(f"Cost: ${trace.total_cost:.3f}")
    print(f"Duration: {trace.duration}s")

Webhook Events

Receive real-time updates:

# Set up webhooks for trace events
webhook_config = {
    "url": "https://your-app.com/webhooks/opswald",
    "events": ["trace.completed", "trace.error", "session.started"],
    "secret": "your-webhook-secret"
}

Dashboard Settings

Team Management

• Invite members with role-based permissions
• Viewer - Read-only access to traces
• Developer - Full debug access + replay
• Admin - User management + billing

Customization

• Dark/light themes
• Timezone settings
• Default view preferences
• Alert thresholds

Data Controls

• Retention period - How long to keep traces (7d to 1y)
• Content filtering - Global PII redaction rules
• Export formats - JSON, CSV, or custom

Getting Help

In-App Support

• Live chat for technical questions
• Documentation search built into the dashboard
• Error explanations with suggested fixes

Community Resources

• Discord community for peer support
• GitHub discussions for feature requests
• Blog tutorials for advanced debugging techniques

The dashboard is designed to make AI agent debugging as intuitive as debugging traditional code. Start with the basic trace view, then explore replay and decision graphs as you need deeper insights.