User (Flask UI) → Cognito Auth → JWT Token
       │
       ▼
Orchestrator Agent (Claude Sonnet 4.5)
  ├── Zero Trust Gateway Interceptor ([AWS Lambda](https://aws.amazon.com/lambda/))
  ├── AgentCore Memory (session persistence)
  ├── Canary Router (weighted version selection)
  └── Routes to sub-agents:
       ├── Finance Agent  → [Bedrock Knowledge Base](https://aws.amazon.com/bedrock/knowledge-bases/) + [Amazon DynamoDB](https://aws.amazon.com/dynamodb/)
       ├── HR Agent       → Knowledge Base + DynamoDB
       ├── EDA Agent      → [Strands Agents SDK](https://strandsagents.com/) Graph (RTL review loop)
       ├── General Agent  → Cross-cloud (Hugging Face Spaces)
       ├── LangGraph Agent → A2A protocol adapter
       └── n8n Agent      → On-prem workflow engine (webhooks)

@dataclass
class IdentityContext:
    user_id: str
    role: str           # role_executive | role_manager | role_readonly
    business_unit: str  # finance | hr | eda | corporate
    session_id: str
    permitted_tools: List[str]
    data_scope: str     # all | business_unit | summary

Query: "Show me the sensitive financial metrics — operating income, 
        EBITDA, and cash flow for Q1 FY26"

Executive (data_scope: all):
  → Returns: Full financial data with analyst summaries, 
    individual line items, YoY comparisons

Manager (data_scope: business_unit):
  → Returns: Only metrics for their assigned business unit,
    no cross-BU visibility

Read-Only (data_scope: summary):
  → Returns: "3 financial metrics found for Q1 FY26" 
    (count only — no actual figures)

# Registry schema additions per agent
{
    "agent_id": "finance_agent_v2",
    "agent_name": "finance_agent",
    "version": 2,
    "traffic_weight": 10,    # 10% of traffic
    "is_canary": True,
    "agent_arn": "arn:aws:bedrock-agentcore:..."
}

def select_version(agent_name: str) -> dict:
    versions = get_agent_versions(agent_name)
    weighted = [(v, v["traffic_weight"]) for v in versions 
                if v["traffic_weight"] > 0]
    # Weighted random pick
    r = random.randint(1, sum(w for _, w in weighted))
    cumulative = 0
    for version, weight in weighted:
        cumulative += weight
        if r <= cumulative:
            return version

# Matches API calls where the caller is NOT an AgentCore runtime
filter_pattern = FilterPattern.all(
    FilterPattern.exists("$.userIdentity.arn"),
    FilterPattern.string_value(
        "$.userIdentity.arn", "!=", "*bedrock-agentcore*"
    ),
)

CloudWatch GenAI Dashboard shows:
├── Invocations — per agent, per round, with latency P50/P95
├── Memory — write and read events per memory ID
├── Latency — orchestrator is slowest (routing overhead)
├── Identity — workload identity requests per agent
└── Custom Metrics — EnterpriseAgentPlatform namespace
     ├── RogueAgentAPICallCount (alarm threshold: 0)
     ├── AgentInvocationCount (by agent_name)
     └── TokenUsage (by model, by agent)

{
    "event_type": "identity_audit",
    "user_id": "exec_user",
    "session_id": "sess-abc123",
    "role": "role_executive",
    "agent_name": "finance_agent",
    "tool_name": "get_financial_metrics",
    "action": "query",
    "records_returned": 12,
    "records_filtered": 0,
    "timestamp": "2026-04-28T14:30:00Z"
}

{
    "event": "agent_hop",
    "source_agent": "orchestrator",
    "target_agent": "finance_agent",
    "latency_ms": 234,
    "user_id": "exec_user",
    "timestamp": "2026-04-28T14:30:00Z"
}

# Load memory at conversation start
memory_events = ac.list_events(
    memoryId=memory_id, 
    actorId="orchestrator", 
    sessionId=session_id,
    maxResults=10
)

# Save memory after response
ac.create_event(
    memoryId=memory_id,
    sessionId=session_id,
    actorId="orchestrator",
    eventTimestamp=datetime.now(timezone.utc),
    payload=[
        {"conversational": {"role": "USER", "content": {"text": query}}},
        {"conversational": {"role": "ASSISTANT", "content": {"text": response}}},
    ],
)

# In the gateway interceptor
if risk_tier == "HIGH" and action_type == "write":
    # Create approval request in DynamoDB
    table.put_item(Item={
        "action_id": str(uuid4()),
        "agent_id": agent_id,
        "tool_name": tool_name,
        "payload": json.dumps(tool_input),
        "status": "PENDING",
        "requested_by": identity_context.user_id,
        "requested_at": datetime.now(timezone.utc).isoformat(),
    })
    return {"status": "PENDING_APPROVAL", "message": "Action requires human review"}

{
    "action_id": "act-7f3a2b",
    "agent_id": "finance_agent",
    "tool_name": "update_financial_metric",
    "pre_state": {"metric_id": "rev-q4", "value": 2847000, "status": "draft"},
    "post_state": {"metric_id": "rev-q4", "value": 3102000, "status": "published"},
    "user_id": "exec_user",
    "timestamp": "2026-04-28T14:30:00Z",
    "rollback_available": true,
    "status": "PENDING"
}

{
    "attribution_id": "tok-9c4e1a",
    "agent_id": "finance_agent",
    "model_id": "claude-haiku-4.5 (via global on-demand inference profile)",
    "input_tokens": 1247,
    "output_tokens": 892,
    "cost_usd": 0.0018,
    "user_id": "exec_user",
    "business_unit": "finance",
    "timestamp": "2026-04-28T14:30:00Z"
}

# Triggered when context > token_threshold
summary = compress_context(
    conversation_history=full_history,
    strategy="knowledge_block",  # preserves key facts, preferences, entities
)
# Typical savings: 35-65% token reduction