Tutorial: Creating New Agents

Every team has unique infrastructure, monitoring systems, and incident response processes. The example agents in our library serve as references and starting points, but the real power of Unpage comes from understanding the agent-building process itself. This tutorial walks you through the steps needed to design and implement your own agents from scratch.

Overview: The Agent Creation Process

Creating a new agent involves six key steps:

Identify your input source - What webhook/alert will trigger the agent?
Write the agent description - How will the router know when to use this agent?
Design the runbook instructions - What should the agent do step-by-step?
Define required tools - What built-in and custom tools does the agent need?
Create custom shell tools - Extend Unpage with your specific commands/scripts
Test and deploy - Validate locally and set up production webhook handling

Let’s walk through each step with a practical example.

Example Scenario: Redis Memory Usage Alerts

For this tutorial, we’ll create an agent that handles Redis memory usage alerts from DataDog. When Redis memory usage exceeds 85%, our agent will:

Check current Redis memory statistics
Identify the largest keys consuming memory
Analyze recent memory growth patterns
Check for memory-intensive operations in Redis logs
Post actionable recommendations to the incident

Step 1: Identify Your Input Source

The first step is understanding what triggers your agent. This could be:

PagerDuty incidents from various monitoring systems
Direct webhooks from DataDog, New Relic, CloudWatch, etc.
GitHub Actions failures or other CI/CD events
Custom application alerts from your own services

For our Redis example, we’ll assume we get alerts that look like:

{
  "incident": {
    "title": "Redis Memory Usage Critical",
    "description": "redis-prod-cluster memory usage: 87.2% (6.1GB/7.0GB)",
    "service": "redis-prod-cluster",
    "status": "triggered"
  }
}

Step 2: Write the Agent Description

The agent description is used by Unpage’s Router to automatically select which agent should handle each incoming alert. Write descriptions that are:

Specific about the alert types this agent handles
Distinguishing to differentiate from other agents
Comprehensive to cover edge cases and variations

Create the agent configuration:

$ unpage agent create redis_memory_alerts

Start with the description in the YAML file that opens:

description: >
  Handle Redis memory usage alerts and high memory consumption issues.
  Use this agent when:
    - The alert mentions Redis, redis-server, or Redis cluster names
    - Memory usage, memory consumption, or OOM (out of memory) is mentioned
    - Redis-specific metrics like used_memory, maxmemory, or evicted_keys are referenced
    - The alert comes from DataDog, CloudWatch, or other monitoring systems monitoring Redis instances

Step 3: Design the Runbook Instructions

The prompt section contains step-by-step instructions for what the agent should do. Think of this as a detailed runbook that a human SRE would follow, but written for an LLM. Structure your instructions clearly:

Use numbered or bulleted steps
Be specific about what information to gather
Include error handling and edge cases
Specify what actions to take based on findings
Include formatting requirements for status updates

prompt: >
  You are a Redis memory analysis specialist. When investigating Redis memory alerts:

  1. Extract the Redis instance/cluster name from the PagerDuty alert
  2. Use `shell_redis_memory_info` to get current memory statistics and configuration
  3. Use `shell_redis_top_keys` to identify the largest keys consuming memory
  4. Use `shell_redis_memory_usage_history` to analyze memory growth patterns over the last 4 hours
  5. Use `search_datadog_logs` to find Redis logs from the last 30 minutes, looking for:
     - Memory-related warnings or errors
     - Large key operations (HSET, SADD with many members)
     - Client connection spikes that might indicate memory leaks
  6. Use `get_resource_with_neighbors` to identify applications connected to this Redis instance
  7. For each connected application, search logs for Redis-related errors or unusual patterns

  Analysis and Response:
  - If memory usage is above 90%: Mark as CRITICAL and recommend immediate action
  - If memory usage is 85-90%: Mark as HIGH and suggest proactive measures
  - If large keys (>100MB) exist: Identify the key patterns and suggest optimization
  - If memory growth is rapid (>10% in 1 hour): Flag as potential memory leak

  Create a comprehensive status update including:
  - Current memory usage percentage and absolute values
  - Top 10 memory-consuming key patterns with sizes
  - Memory growth rate over the last 4 hours
  - Any concerning log patterns or errors
  - Connected applications that might be causing issues
  - Specific recommended actions (key cleanup, configuration changes, scaling)

  Post findings using `pagerduty_post_status_update` with priority based on severity analysis.

Step 4: Define Required Tools

List all the tools your agent needs in the tools section. These include:

Built-in tools from Unpage plugins (DataDog, PagerDuty, AWS, etc.)
Custom shell commands you’ll create for specific operations
Wildcards for groups of related tools

tools:
  - "shell_redis_memory_info"
  - "shell_redis_top_keys"
  - "shell_redis_memory_usage_history"
  - "search_datadog_logs"
  - "get_resource_with_neighbors"
  - "pagerduty_post_status_update"

To see all available built-in tools:

$ unpage mcp tools list

Your Agent will only have access to the tools you explicitly give it permission to call.

Step 5: Create Custom Shell Tools

You can always extend Unpage with custom shell commands to interact with your specific infrastructure. These commands can:

Execute Redis CLI commands against your instances
Run custom scripts or database queries
Call internal APIs or tools
Parse and format data for the agent

Edit your Unpage configuration (~/.unpage/profiles/default/config.yaml) to add the custom commands:

plugins:
  # ... existing plugins
  shell:
    enabled: true
    settings:
      commands:
        - handle: redis_memory_info
          description: Get comprehensive Redis memory statistics and configuration
          command: |
            redis-cli -h {redis_host} -p {redis_port} --raw INFO memory &&
            echo "---CONFIG---" &&
            redis-cli -h {redis_host} -p {redis_port} CONFIG GET maxmemory* &&
            redis-cli -h {redis_host} -p {redis_port} CONFIG GET save
          args:
            redis_host: The Redis server hostname or IP address
            redis_port: The Redis server port (default 6379)

        - handle: redis_top_keys
          description: Identify the largest keys in Redis by memory usage
          command: |
            redis-cli -h {redis_host} -p {redis_port} --latency-history -i 1 > /dev/null 2>&1 &
            LATENCY_PID=$!
            redis-cli -h {redis_host} -p {redis_port} --bigkeys --i 0.01
            kill $LATENCY_PID 2>/dev/null || true
          args:
            redis_host: The Redis server hostname or IP address
            redis_port: The Redis server port (default 6379)

        - handle: redis_memory_usage_history
          description: Get Redis memory usage metrics from the last 4 hours via DataDog API
          command: |
            curl -X GET "https://api.datadoghq.com/api/v1/query" \
              -H "Content-Type: application/json" \
              -H "DD-API-KEY: ${DATADOG_API_KEY}" \
              -H "DD-APPLICATION-KEY: ${DATADOG_APP_KEY}" \
              -G \
              --data-urlencode "query=avg:redis.info.memory.used_memory{host:{redis_host}}" \
              --data-urlencode "from=$(date -d '4 hours ago' +%s)" \
              --data-urlencode "to=$(date +%s)"
          args:
            redis_host: The Redis server hostname to query metrics for

Shell Command Best Practices

When creating shell commands:

Include error handling with 2>/dev/null || echo "Command failed"
Use environment variables for API keys and credentials
Chain commands with && for sequential execution
Parse output to provide clean, structured data
Add timeouts for potentially long-running operations
Document required permissions and dependencies

Step 6: Test and Deploy

Local Testing

Test your agent with sample data before deploying:

# Test with a sample alert payload
$ echo '{"incident": {"title": "Redis Memory Usage Critical", "description": "redis-prod-cluster memory usage: 87.2%"}}' | unpage agent run redis_memory_alerts

# Test with a PagerDuty incident ID
$ unpage agent run redis_memory_alerts --pagerduty-incident PXXXXX

Test Routing

Verify the router selects your agent correctly:

# Test routing decision
$ unpage agent route '{"incident": {"title": "Redis Memory Critical"}}'

# Debug routing with detailed explanation
$ unpage agent route --debug '{"incident": {"title": "Redis Memory Critical"}}'

Production Deployment

Set up webhook handling for production alerts:

# Local webhook server for testing
$ unpage agent serve

# Public webhook with ngrok tunnel
$ unpage agent serve --tunnel --ngrok-token YOUR_NGROK_TOKEN

# Production deployment (typically with reverse proxy)
$ unpage agent serve --host 0.0.0.0 --port 8000

Configure your monitoring system (PagerDuty, DataDog, etc.) to send webhooks to:

Local testing: http://localhost:8000/webhook
Ngrok tunnel: https://your-tunnel.ngrok.io/webhook
Production: https://your-domain.com/webhook

Advanced Agent Patterns

Multi-Step Analysis Agents

For complex scenarios, break analysis into phases:

prompt: >
  Phase 1 - Data Collection:
  - Gather all relevant metrics and logs
  - Verify the scope of the issue

  Phase 2 - Root Cause Analysis:
  - Correlate data to identify potential causes
  - Rule out common false positives

  Phase 3 - Impact Assessment:
  - Determine affected services and users
  - Estimate business impact

  Phase 4 - Response and Communication:
  - Post detailed findings with evidence
  - Recommend specific remediation steps
  - Set appropriate incident priority

Conditional Logic Agents

Use conditional prompts for different scenarios:

prompt: >
  Analyze the alert and determine the scenario:

  If memory usage > 95%:
    - Execute emergency memory cleanup procedures
    - Post CRITICAL update with immediate actions

  If memory growth rate > 20% per hour:
    - Focus on identifying memory leaks
    - Examine recent deployments and configuration changes

  If evicted_keys metric is increasing:
    - Analyze key eviction patterns
    - Recommend maxmemory policy adjustments

  Otherwise:
    - Perform standard memory analysis
    - Post standard monitoring recommendations

Integration with External Systems

Agents can interact with any system your shell commands can reach:

tools:
  - "shell_slack_notify_team"
  - "shell_create_jira_ticket"
  - "shell_trigger_runbook_automation"
  - "shell_update_status_page"

Debugging and Iteration

Monitoring Agent Performance

Use Unpage’s built-in tracing to monitor agent execution:

# Start MLflow tracking server
$ unpage mlflow serve

# Run agent with tracing enabled
$ env MLFLOW_TRACKING_URI=http://127.0.0.1:5566 unpage agent run redis_memory_alerts @test_alert.json

View execution traces at

http://127.0.0.1:5566/#/experiments/1?searchFilter=&orderByKey=attributes.start_time&orderByAsc=false&startTime=ALL&lifecycleFilter=Active&modelVersionFilter=All+Runs&datasetsFilter=W10%3D&compareRunsMode=TRACES

to see:

Tool usage patterns
Execution timing
Error rates and types
Agent decision flows

More details are in the Debugging with MLflow tracing guide.

Best Practices Summary

Start simple - Begin with basic analysis, then add complexity
Test thoroughly - Use various input scenarios and edge cases
Handle errors gracefully - Include fallbacks for failed commands
Be specific in descriptions - Help the router make correct decisions
Document dependencies - Note required tools, permissions, and environment setup
Iterate based on results - Refine prompts based on real incident responses
Monitor and improve - Use tracing data to optimize agent performance

Conclusion

Creating effective Unpage agents transforms reactive incident response into proactive, automated analysis. By following this systematic approach you can build agents that not only save time during incidents but also provide deeper insights into your infrastructure than manual investigation alone. The key is starting with one well-defined use case, perfecting it through testing and iteration, then expanding to cover additional scenarios as you gain experience with the platform. Remember: the best agents are those that encode your team’s operational knowledge and decision-making processes, making your entire team more effective at infrastructure management.

About Unpage

Core Concepts

Example Agents

Plugins

How-To Guides

Command Reference

Tutorial: Creating New Agents

Overview: The Agent Creation Process

Example Scenario: Redis Memory Usage Alerts

Step 1: Identify Your Input Source

Step 2: Write the Agent Description

Step 3: Design the Runbook Instructions

Step 4: Define Required Tools

Step 5: Create Custom Shell Tools

Shell Command Best Practices

Step 6: Test and Deploy

Local Testing

Test Routing

Production Deployment

Advanced Agent Patterns

Multi-Step Analysis Agents

Conditional Logic Agents

Integration with External Systems

Debugging and Iteration

Monitoring Agent Performance

Best Practices Summary

Conclusion

​Overview: The Agent Creation Process

​Example Scenario: Redis Memory Usage Alerts

​Step 1: Identify Your Input Source

​Step 2: Write the Agent Description

​Step 3: Design the Runbook Instructions

​Step 4: Define Required Tools

​Step 5: Create Custom Shell Tools

​Shell Command Best Practices

​Step 6: Test and Deploy

​Local Testing

​Test Routing

​Production Deployment

​Advanced Agent Patterns

​Multi-Step Analysis Agents

​Conditional Logic Agents

​Integration with External Systems

​Debugging and Iteration

​Monitoring Agent Performance

​Best Practices Summary

​Conclusion

Overview: The Agent Creation Process

Example Scenario: Redis Memory Usage Alerts

Step 1: Identify Your Input Source

Step 2: Write the Agent Description

Step 3: Design the Runbook Instructions

Step 4: Define Required Tools

Step 5: Create Custom Shell Tools

Shell Command Best Practices

Step 6: Test and Deploy

Local Testing

Test Routing

Production Deployment

Advanced Agent Patterns

Multi-Step Analysis Agents

Conditional Logic Agents

Integration with External Systems

Debugging and Iteration

Monitoring Agent Performance

Best Practices Summary

Conclusion