Deploy Multi-Agent Service

When to Use

• Deploy multi-agent orchestration infrastructure
• Configure supervisor agent and worker agent topology
• Set up handoff protocol between agents
• Configure shared context store for inter-agent state
• Deploy timeout policies and circuit breakers

Prerequisites

1. Azure CLI authenticated: az account show
2. Bicep CLI: az bicep version
3. Azure OpenAI with multiple model deployments
4. Azure Container Apps or AKS for agent hosting
5. Azure Redis or Cosmos DB for shared state

Step 1: Validate Infrastructure

az bicep lint -f infra/main.bicep
az bicep build -f infra/main.bicep

Verify resources:

• Azure OpenAI (multiple deployments: gpt-4o for supervisor, gpt-4o-mini for workers)
• Azure Container Apps (one container per agent role)
• Azure Redis Cache (shared context store, agent state)
• Azure Service Bus (agent-to-agent message queue)
• Azure Monitor (distributed tracing across agents)

Step 2: Deploy Azure Resources

az deployment group create \
  --resource-group $RESOURCE_GROUP \
  --template-file infra/main.bicep \
  --parameters infra/parameters.json

Step 3: Configure Agent Topology

Define the agent graph in config/agents.json:

Step 4: Configure Handoff Protocol

# Handoff message format between agents
handoff = {
    "from_agent": "supervisor",
    "to_agent": "researcher",
    "task": "Find recent sales data for Q4",
    "context": { "shared_state_key": "task-123" },
    "deadline": "2026-01-01T00:00:30Z",  # 30s timeout
    "retry_policy": { "max_retries": 2, "backoff": "exponential" }
}

Step 5: Configure Shared Context Store

• Use Azure Redis for fast inter-agent state sharing
• Key pattern: task:{task_id}:agent:{agent_name}:state
• TTL: 1 hour (auto-cleanup completed tasks)
• Max context size: 4KB per agent per task

Step 6: Configure Safety Controls

• Loop detection: Max 10 handoffs per task (prevent infinite delegation loops)
• Timeout cascade: If worker times out, supervisor retries with different worker
• Circuit breaker: If agent fails 3 consecutive tasks, remove from topology
• Cost cap: Max $0.50 per task across all agents
• Human escalation: After 2 failed attempts, escalate to human

Step 7: Smoke Test

python scripts/test_topology.py --task "Summarize Q4 sales and write a report"
python scripts/test_handoff.py --from supervisor --to researcher
python scripts/test_loop_detection.py --max-hops 15  # Should trigger at 10
python scripts/test_timeout.py --agent researcher --delay 25s  # Should timeout at 20s

Post-Deployment Verification

• [ ] Supervisor correctly decomposes tasks into subtasks
• [ ] Handoffs between agents complete within timeout
• [ ] Shared context store accessible by all agents
• [ ] Loop detection triggers at configured threshold
• [ ] Circuit breaker activates on repeated failures
• [ ] Distributed tracing shows full agent chain
• [ ] Cost per task within budget cap

Troubleshooting