deploy-capability-agents

name: deploy-capability-agents description: Deploys optional A2A capability agents (Knowledge Base, CRM) that extend the voice agent with new skills. Explains the A2A architecture, enables the capability registry, deploys agent stacks, and verifies discovery. Use after the core deployment and Daily setup are complete.

Deploy Capability Agents

You are guiding the user through extending their voice agent with A2A (Agent-to-Agent) capability agents. This is optional — the voice agent works without them, but they add powerful features.

When This Skill Activates

• User says "deploy agents", "add KB", "add CRM", "capability agents"
• User finished /configure-daily and wants to extend the agent
• User asks about knowledge base, RAG, customer lookup, or A2A

Prerequisites

The core deployment must be complete:

• /deploy-cloud-api or /deploy-sagemaker has been run
• /configure-daily has been run (phone number active)
• The user has tested the basic agent (time check, hangup)

What's Running Now

Before deploying anything, set context for the user:

Current voice agent capabilities:
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
  • Natural conversation — powered by Claude on Bedrock
  • get_current_time    — "What time is it?"
  • hangup_call         — "Goodbye" / ask to end the call

Tool calling is enabled by default. The agent can hold conversations
and use the two built-in tools above. No other tools are active yet.

How Capability Agents Work

Explain the architecture before deploying:

A2A Architecture:
━━━━━━━━━━━━━━━━
                    ┌─────────────────┐
                    │   Voice Agent   │
                    │   (ECS hub)     │
                    └────────┬────────┘
                             │ polls every 30s
                    ┌────────▼────────┐
                    │    CloudMap     │
                    │  (discovery)    │
                    └──┬──────────┬──┘
                       │          │
              ┌────────▼──┐  ┌───▼────────┐
              │  KB Agent  │  │ CRM Agent  │
              │ (ECS spoke)│  │(ECS spoke) │
              └────────────┘  └────────────┘

Key points:

• Each capability agent runs as a separate ECS container (its own task definition, scaling, and logs)
• Agents register themselves in CloudMap under a shared namespace
• The voice agent polls CloudMap every 30 seconds for new agents
• When discovered, the agent's skills are automatically registered as LLM tools — no pipeline code changes needed
• If an agent goes away, its tools are automatically deregistered after a grace period

What To Do

Phase 1: Check Current State

Verify the core deployment is healthy and check whether agents are already deployed:

# Core health
aws ssm get-parameter --name "/voice-agent/ecs/cluster-arn" --query 'Parameter.Value' --output text 2>/dev/null

# Capability registry status
aws ssm get-parameter --name "/voice-agent/config/enable-capability-registry" --query 'Parameter.Value' --output text 2>/dev/null || echo "not-set (defaults to false)"

# Agent stacks
aws cloudformation describe-stacks --stack-name VoiceAgentKbAgent --query 'Stacks[0].StackStatus' --output text 2>/dev/null || echo "not-deployed"
aws cloudformation describe-stacks --stack-name VoiceAgentCrmAgent --query 'Stacks[0].StackStatus' --output text 2>/dev/null || echo "not-deployed"

Phase 2: Choose Agents

Present the available agents and let the user choose:

Ask: "Which agents would you like to deploy? You can deploy one or both."

Phase 3: Enable the Capability Registry

This is required once before deploying any agents. Skip if already enabled.

aws ssm put-parameter \
  --name "/voice-agent/config/enable-capability-registry" \
  --value "true" --type String --overwrite

The voice agent reads this flag at the start of each call. No container restart is needed — the next incoming call will start the CloudMap polling loop.

Phase 4: Deploy Selected Agents

Get confirmation before deploying. Show the account ID and region.

Knowledge Base Agent

USE_CLOUD_APIS=true npx cdk deploy VoiceAgentKnowledgeBase VoiceAgentKbAgent --require-approval never

This deploys:

• VoiceAgentKnowledgeBase — Bedrock Knowledge Base + S3 bucket with sample documents from resources/
• VoiceAgentKbAgent — ECS Fargate container that exposes a search_knowledge_base skill via A2A

Takes 5-8 minutes. The KB agent uses a direct tool executor (bypasses inner LLM) for low-latency queries (300ms).

CRM Agent

USE_CLOUD_APIS=true npx cdk deploy VoiceAgentCRM VoiceAgentCrmAgent --require-approval never

This deploys:

• VoiceAgentCRM — DynamoDB tables (empty) + CRM API Lambda for customers and cases
• VoiceAgentCrmAgent — ECS Fargate container that exposes 5 CRM tools via A2A (customer lookup, case creation, case update, case listing, case detail)

Takes ~5-8 minutes.

Seed the CRM with demo data — the DynamoDB tables are created empty. The CRM API has a /admin/seed endpoint that loads demo customers and cases:

CRM_URL=$(aws ssm get-parameter --name "/voice-agent/crm/api-url" --query 'Parameter.Value' --output text)
curl -s -X POST "$CRM_URL/admin/seed" | python3 -m json.tool

Expected response: {"message": "Demo data seeded successfully", "customers_seeded": 3, "cases_seeded": 2}

The seed data includes:

There are also 2 demo support cases (a billing dispute for John Smith and a service outage for Michael Chen).

Without seeding, all CRM lookups return "customer not found." This step is required for testing.

To reset and re-seed later: DELETE /admin/reset followed by POST /admin/seed.

Phase 5: Verify Agent Discovery

After deployment, wait ~1 minute for the agents to register in CloudMap and be discovered. Then verify:

# Check CloudMap namespace for registered services
NAMESPACE=$(aws ssm get-parameter --name "/voice-agent/ecs/cloudmap-namespace" --query 'Parameter.Value' --output text 2>/dev/null)
if [ -n "$NAMESPACE" ]; then
  NAMESPACE_ID=$(aws servicediscovery list-namespaces --query "Namespaces[?Name=='$NAMESPACE'].Id" --output text)
  aws servicediscovery list-services --filters "Name=NAMESPACE_ID,Values=$NAMESPACE_ID,Condition=EQ" \
    --query 'Services[].Name' --output text
fi

You should see the agent service names listed. The voice agent will discover them on the next CloudMap poll (within 30 seconds).

Phase 6: Test

Prompt the user to make a test call:

Capability agents deployed:
━━━━━━━━━━━━━━━━━━━━━━━━━━
  [status] Knowledge Base — "What's your return policy?"
  [status] CRM           — "Look up the account for 555-0100"

Call [phone number] and try one of the prompts above.
The voice agent discovers new agents automatically — no restart needed.

If the CRM agent was deployed, confirm the seed data is loaded:

CRM_URL=$(aws ssm get-parameter --name "/voice-agent/crm/api-url" --query 'Parameter.Value' --output text)
curl -s "$CRM_URL/customers" | python3 -c "
import json, sys
data = json.load(sys.stdin)
if not data:
    print('WARNING: No customers found. Run POST /admin/seed to load demo data.')
else:
    print(f'{len(data)} customer(s) loaded:')
    for c in data:
        print(f'  {c.get(\"name\", \"?\")} — {c.get(\"phone\", \"?\")}')
"

If a tool isn't responding on the first call, remind the user:

• The registry polls every 30 seconds — wait a moment and try again
• Check ECS console to confirm the agent container is running and healthy
• Check CloudWatch logs for the agent container if it's not registering

Phase 7: Report

Capability Agent Deployment Complete:
✅ Capability registry enabled
[status] Knowledge Base agent — deployed / skipped
[status] CRM agent — deployed / skipped
✅ Agents discoverable via CloudMap

Your voice agent now has [N] additional tools available.
Run /verify-deployment to confirm full system health.

Remind user to use the destroy-project skill to tear down all resources when done.