Delegating to CDN Agent

Delegating to CDN Agent

Core Principle

Never handle Fastly CDN operations directly. Always delegate to a specialized sub-agent to keep your context clean and costs low.

Recognition Patterns

Delegate when user says:

• "query fastly for..."
• "list fastly services"
• "show cdn configuration"
• "purge cache"
• "describe fastly service"
• "list backends"
• "show VCL"
• "clear cdn cache"
• Any mention of: fastly, cdn, cache, vcl, backends, purge, origin servers, varnish

ExampleJobInc-Specific Triggers

Also delegate when user mentions these ExampleJobInc domains:

• "pi.examplejobinc-cdn.com"
• "images.examplejobinc.com"
• "hls.examplejobinc.com"
• "assets.examplejobinc.com"
• "embed.examplejobinc.com"
• "crushinator" (image processing backend)

How to Delegate

Use the Task tool with a specialized prompt:

Task(
  subagent_type: "general-purpose",
  description: "Query Fastly CDN",
  prompt: "<full agent instructions from AGENT-INSTRUCTIONS.md>"
)

Agent Prompt Template

When delegating, include:

1. The complete agent instructions (see AGENT-INSTRUCTIONS.md)
2. The user's specific request
3. Clear output format requirements

Example:

You are a Fastly CDN specialist. Your job is to query Fastly using shell wrappers and return clean results.

<AGENT INSTRUCTIONS HERE>

USER REQUEST: List all Fastly services

Return a clean summary with:
- Service names
- Service IDs
- Service types
- Last updated dates

After Agent Returns

1. Present results cleanly to user
2. Offer follow-up if relevant (e.g., "Would you like to see backends for pi.examplejobinc-cdn.com?")
3. Don't expose mechanics (API calls, tokens, CLI commands) to user

Benefits

• ✅ Main context stays clean
• ✅ Cheaper queries (sub-agent uses less expensive model)
• ✅ Specialized knowledge isolated
• ✅ Scalable pattern for other CDN operations

Example Flow

User: "list fastly services"

Main Assistant: [Recognizes Fastly query]
              → Invokes Task tool with agent instructions
              → Agent runs cdn-services wrapper
              → Agent returns formatted results

Main Assistant: "Found 8 Fastly services:

                1. pi.examplejobinc-cdn.com - Image processing CDN
                2. images.examplejobinc.com - Static images
                3. hls.examplejobinc.com - Video streaming
                ...

                Would you like details on any of these?"

Common Query Types

Service Management:

• "list fastly services"
• "describe pi.examplejobinc-cdn.com service"
• "show cdn configuration"

Backend Investigation:

• "list backends for pi.examplejobinc-cdn.com"
• "show origin servers"
• "describe crushinator backend"

VCL Configuration:

• "list VCL configurations"
• "show VCL content for version 42"
• "view custom VCL"

Cache Management:

• "purge all cache for pi.examplejobinc-cdn.com"
• "clear cdn cache"
• "invalidate surrogate key image-cache"
• "purge specific URL"

Red Flags

DON'T:

• ❌ Try to run fastly-* scripts yourself
• ❌ Attempt Fastly API calls directly
• ❌ Load detailed Fastly CLI knowledge
• ❌ Handle authentication directly
• ❌ Parse complex CDN configurations

DO:

• ✅ Immediately delegate on Fastly keywords
• ✅ Trust the sub-agent's results
• ✅ Present clean summaries to user
• ✅ Suggest relevant follow-up queries
• ✅ Warn about cache purge impact

Cache Purge Safety

When user requests cache purging:

1. Always delegate to sub-agent (never do yourself)
2. Confirm service before agent purges
3. Warn about impact of full cache purges
4. Prefer specific purges (URL/key) over --all

The sub-agent handles this, but you should reinforce warnings in your response.

ExampleJobInc Infrastructure Context

ExampleJobInc uses Fastly for:

• Image CDN (pi.examplejobinc-cdn.com) - Crushinator-powered image processing
• Video CDN (hls.examplejobinc.com) - HLS streaming
• Static assets (images.examplejobinc.com, assets.examplejobinc.com)
• Website CDN (www.examplejobinc.com)

When user mentions investigating these systems, delegate to Fastly agent.

Version History

• 1.0.0 (2025-10-15): Initial delegation skill created to reduce context pollution