AgentCore
AgentCore handles tool execution for the AI Assistant, powered by Amazon Bedrock AgentCore Runtime and Gateway. Unlike v1's single-EC2 embedded approach, v2 splits everything out serverless: Runtime + 9 section Gateways + Memory + Code Interpreter.
8 AWS-domain Gateways + external-obs (external observability) = 9 routed sections · 160 MCP tools across the full catalog (the 27 Lambda targets — vendor-hosted mcpServer targets excluded) · 30 Lambda slices (21 gated on agentcore_enabled, 9 on integrations_enabled, both off by default) run serverless. The classifier routes each question to 1-3 routes and calls them in parallel, then synthesizes. → Why AWSops
Architecture
AI Routing Flow
Deployment Requirements
| Item | Requirement |
|---|---|
| Docker | arm64 required (docker buildx --platform linux/arm64 --load; make agentcore builds and pushes) |
| agent.py | Gateway URLs are injected via the GATEWAYS_JSON env (no per-account hardcoding) |
| Code Interpreter / Memory | No hyphens in name, underscores only |
| Memory Store | Max 365-day retention (eventExpiryDuration) |
| Config source of truth | SSM /ops/awsops-v2/agentcore/{runtime_arn,interpreter_id,memory_id} — written by provision.py, read by the web BFF at runtime (never exposed in the UI) |
| Runtime updates | Re-running the idempotent provisioner (scripts/v2/agentcore/provision.py) applies changes — a just-created Gateway not yet READY can make the first target creation fail; re-run resolves it |
AgentCore Runtime
Configuration
| Item | Description |
|---|---|
| Engine | Strands Agent Framework |
| Container | Docker arm64 (stored in ECR, via make agentcore) |
| Execution Environment | AgentCore managed service (Bedrock AgentCore Runtime) |
| Model | Claude Haiku 4.5 (ADR-038 routing classifier) / the Runtime itself only ever runs Sonnet 4.6 (hardcoded in agent/agent.py). Opus 4.8 isn't used by the Runtime — it's an optional Deep-tier choice in the separate AI Diagnosis (async worker) feature |
Status
- READY: Operating normally
- CREATING: Being created
- UPDATING: Being updated
- FAILED: Error state
Gateway Details
v2 has 8 AWS-domain Gateways (awsops-v2-{network,container,data,security,cost,monitoring,iac,ops}-gateway) plus external-obs — a routed section hosting external-observability/integration connectors, chat-aliased to the observability routing key. Tool counts below are from the full catalog (scripts/v2/agentcore/catalog.py); activation is gated behind the agentcore_enabled/integrations_enabled flags (off by default on new installs). The live environment has the fleet fully deployed — all 9 gateways hold READY MCP targets and all 16 chat section keys are active (2026-08-02).
Network Gateway (17 tools)
Provides VPC, ENI, Reachability, Flow Logs, TGW, VPN, and Network Firewall tools.
| Category | Tools |
|---|---|
| flow-monitor | query_flow_logs |
| reachability-read | check_reachability |
| network-mcp | get_path_trace_methodology, find_ip_address, get_eni_details, list_vpcs, get_vpc_network_details, get_vpc_flow_logs, describe_network, list_transit_gateways, get_tgw_details, get_tgw_routes, get_all_tgw_routes, list_tgw_peerings, list_vpn_connections, list_network_firewalls, get_firewall_rules |
Container Gateway (19 tools)
Provides EKS, ECS, and Istio service-mesh tools.
| Category | Tools |
|---|---|
| istio-read | mesh_overview, list_virtual_services, list_destination_rules, list_istio_gateways, list_service_entries, list_authorization_policies, list_peer_authentications |
| eks-mcp | list_eks_clusters, get_eks_vpc_config, get_eks_insights, get_cloudwatch_logs, get_cloudwatch_metrics, get_eks_metrics_guidance, get_policies_for_role, search_eks_troubleshoot_guide, generate_app_manifest |
| ecs-mcp | ecs_resource_management, ecs_troubleshooting_tool, wait_for_service_ready |
IaC Gateway (12 tools)
Provides Infrastructure as Code tools.
| Category | Tools |
|---|---|
| iac-mcp | validate_cloudformation_template, check_cloudformation_template_compliance, troubleshoot_cloudformation_deployment, search_cdk_documentation, search_cloudformation_documentation, cdk_best_practices, read_iac_documentation_page |
| terraform-mcp | SearchAwsProviderDocs, SearchAwsccProviderDocs, SearchSpecificAwsIaModules, SearchUserProvidedModule, terraform_best_practices |
Data Gateway (24 tools)
Provides AWS database and streaming service tools.
| Category | Tools |
|---|---|
| rds-mcp | list_db_instances, list_db_clusters, describe_db_instance, describe_db_cluster, execute_sql, list_snapshots |
| dynamodb-mcp | list_tables, describe_table, query_table, get_item, dynamodb_data_modeling, compute_performances_and_costs |
| msk-mcp | list_clusters, get_cluster_info, get_configuration_info, get_bootstrap_brokers, list_nodes, msk_best_practices |
| valkey-mcp | list_cache_clusters, describe_cache_cluster, list_replication_groups, describe_replication_group, list_serverless_caches, elasticache_best_practices |
Security Gateway (14 tools)
Provides IAM and security analysis tools. (Deployed in P1f.)
| Tool | Description |
|---|---|
list_users / get_user | List/detail IAM users |
list_roles / get_role_details | List/detail IAM roles |
list_groups / get_group | List/detail IAM groups |
list_policies | List policies |
list_user_policies / list_role_policies | List user/role policies |
get_user_policy / get_role_policy | User/role inline policy |
list_access_keys | List Access Keys |
simulate_principal_policy | Policy simulation |
get_account_security_summary | Account security summary |
Monitoring Gateway (36 tools)
Provides CloudWatch and CloudTrail (AWS-native), plus OpenSearch and the Loki/Tempo/Mimir observability stack (Prometheus and ClickHouse are owned by External-Obs — ADR-004).
| Category | Tools |
|---|---|
| cloudwatch-mcp (11) | Metrics/alarms/log-insights queries |
| cloudtrail-mcp (5) | lookup_events, list_event_data_stores, lake_query, get_query_status, get_query_results |
| opensearch-mcp (4) | OpenSearch domain/index queries |
loki-mcp / tempo-mcp / mimir-mcp (loki 5 · tempo 5 · mimir 6, integrations_enabled) | LogQL/TraceQL queries — Loki/Tempo/Mimir stay on this Gateway (ADR-004; Prometheus·ClickHouse moved to External-Obs) |
Cost Gateway (14 tools)
Provides cost analysis, forecasting, and FinOps tools.
| Category | Tools |
|---|---|
| cost-mcp (9) | get_today_date, get_cost_and_usage, get_cost_and_usage_comparisons, get_cost_comparison_drivers, get_cost_forecast, get_dimension_values, get_tag_values, get_pricing, list_budgets |
| finops-mcp (5) | Compute Optimizer rightsizing, RI/SP recommendations, Cost Optimization Hub, Trusted Advisor |
Ops Gateway (11 tools)
Provides AWS documentation search, operations helpers, and inventory query tools.
| Category | Tools |
|---|---|
| core-helpers | prompt_understanding, suggest_aws_commands |
| inventory-read | find_unused_resources, get_topology, query_inventory, inventory_summary |
| aws-knowledge | search_documentation, read_documentation, recommend, list_regions, get_regional_availability |
External-Obs (13 tools, routing key: observability)
The 9th routed section, hosting external-observability/integration connectors (ADR-004, amended 2026-06-24). The catalog defines notion-mcp (3 tools) here, gated on integrations_enabled (off by default). The Prometheus (6 tools) and ClickHouse (4 tools) connector targets also live on this section (catalog.py — prometheus-mcp-target/clickhouse-mcp-target with gateway=external-obs).
Code Interpreter
Provides a sandbox environment for Python code execution.
Features
- Isolated environment: Secure Python execution
- Data analysis: Library support for pandas, numpy, etc.
- Visualization: Chart generation with matplotlib, plotly, etc.
- File processing: Data parsing for JSON, CSV, etc.
Usage Examples
"Visualize AWS cost data as a monthly trend chart"
"Parse this JSON data and calculate statistics"
Routing Display (AI Assistant)
Instead of v1's separate "AgentCore" dashboard page (call statistics, config lookup), v2 shows routing information inline inside the AI Assistant chat screen.
- Each answer shows a badge for which section (Gateway) handled it.
- Answers synthesized from multiple domains in parallel show "via" chips, one per contributing Gateway (e.g.
multi:network+data). - Up to two alternate-route chips let you re-ask via a different route.
- The chat rail lists recent conversation threads (no full-text search yet).
The AgentCore Runtime ARN, Memory ID, and similar config values live only in SSM and are never exposed in the UI (operators check via terraform output / SSM).
Known Limitations
| Item | Limitation |
|---|---|
| Docker architecture | arm64 required |
| Code Interpreter / Memory name | No hyphens, underscores only |
| Conversation history retention | Maximum 365 days |
| AgentCore response | Returns final text only (tool inference streamed with a typing effect) |
| Tool activation flags | The 30 slices are gated on agentcore_enabled (21) / integrations_enabled (9) — the live environment has the fleet fully deployed (2026-08-02); new installs default to off |
| Vendor mcpServer targets | For external-obs's vendor-hosted mcpServer targets (Datadog · Dynatrace · New Relic), capability=read is not enforced at the protocol level (ADR-017) — the read-only guarantee applies to Lambda targets; unset endpoints SKIP. At runtime a fail-closed tool allowlist (OFFICIAL_MCP_TOOL_ALLOWLIST_JSON — only catalog-transcribed read tools pass, agent/agent.py) compensates |
Next Steps
- AI Assistant - Using AI features
- Dashboard - Return to dashboard