| Agent Registry |
No agent-as-artifact concept (versioning, publishing, rollback). Agents are runtime constructs, not stored artifacts. |
| Multi-Agent Orchestration |
No agent coordination, scheduling, recursion-depth limits, or agent profile management |
| Agent Kill-Switch Automation |
Manual suspend/disable exists; no policy-driven automatic global termination on anomaly |
| Prompt Registry |
No centralized prompts with versioning, tagging, or Dev/Staging/Prod environments |
| Prompt Evaluation Framework |
No continuous eval against test sets / scoring rubrics |
| Prompt Drift Analysis |
No baseline tracking for system-prompt modifications |
| Human-in-the-Loop Workflows |
No pause-and-approve flow for Tier-3 tool calls |
| Prompt Playground |
No interactive prompt-testing UI with code-snippet generation |
| Prompt Compression |
No token-reduction step before forwarding |
| A/B Testing / Canary / Blue-Green |
No traffic-splitting framework for model/agent comparison |
| Model Drift Detection |
No quality-degradation monitoring (refusal-rate baseline tracking, etc.) |
| Hallucination / Faithfulness Monitoring |
No automated faithfulness scoring |
| RAG Quality Metrics |
No Context Precision / Recall / Faithfulness scoring per request |
| Context Window Visualization |
Token counts captured; no “lost in the middle” inefficiency detection |
| FOCUS-Compliant Cost Export |
Cost tracking exists but not FinOps FOCUS schema |
| 4-Tier Consumption Reporting (Org / Team / User / Model) |
4-tier budget enforcement ships; the matching report deck is partial |
| Unit Economics + Forecasting |
No predictive cost modelling |
| Cost Anomaly Alerting |
Threshold-based budget alerts exist; no statistical anomaly detection (e.g., “> 3× normal rate”) |
| User Feedback System |
No thumbs-up/down or star rating correlated with trace_id |
| Throttle-Event Visibility UI |
Events emit; no dedicated dashboard view of when/why an agent capability was throttled |
| Topology / Lineage Visualization |
Knowledge-graph viz exists; no Agent → MCP → Data-Source topology map |
| Config Drift Monitoring |
Versioned configs exist; no automated drift alerting against a desired-state baseline |
| Configuration Validation Tool (vendor-spec) |
Schema validation exists; no separate “validate config matches vendor spec” CLI |
| Declarative Policy Engine |
No OPA/Rego or Google-CEL policy language |
| Guardrail Versioning + Rollback |
Configurable + auditable; not formally versioned with rollback |
| Per-Agent Tool Whitelists |
Server-level ACLs only; cannot whitelist tools per individual agent identity |
| Virtual MCP Servers |
Cannot construct virtual MCP servers by curating tool subsets from base servers |
| HashiCorp Vault / AWS Secrets Manager / Azure Key Vault / CyberArk |
Secrets encrypted in DB; no external-vault native integration |
| SPIFFE/SPIRE Workload Identity |
No hardware-attestation-based workload identity |
| NHI Governance (RFC 8705 cert-bound tokens) |
No non-human-identity registration, permission-intersection model, or cert-bound token enforcement |
| Okta as NHI Anchor |
Okta SSO supported for human users; no NHI registration anchored to an Okta-owning human |
| Acryl DataHub Lineage |
No native data-lineage integration (Kafka feed enables external reconstruction) |
| A2A Protocol (advanced) |
Core A2A shipped in the 2.0.64–2.0.105 window — agent registry, agent-card discovery, JSON-RPC 2.0 dispatch, X-A2A-Key + Bearer JWT auth, tenant scoping (§42.11, §13). Deferred: mTLS workload identity, RFC 8693 token exchange / on-behalf-of, push-notification webhooks, external-agent egress proxy. |
| LDAP Group-Based Org Routing |
Schema reserves signalType='group'; only signalType='domain' shipped |
| Anthropic-SDK Cross-Protocol Translation |
/v1/messages is currently a native Anthropic-only passthrough endpoint — does not translate to OpenAI / Bedrock / Vertex upstream providers in v2.0.105. Workaround: use the OpenAI SDK with the same upstream (the Anthropic adapter handles OpenAI→Anthropic translation in the other direction). See §3.6. |
| Bedrock Converse API as Inbound Endpoint |
The raw AWS Bedrock Converse API wire format is not an inbound endpoint. Bedrock is fully supported as an upstream via OpenAI / Vertex / Azure / Cohere SDKs — switch the app to one of those and AgentWatch routes to Bedrock for you. See §3.6. |
| Adversarial / Red-Team CI Framework |
Guardrails catch prompt-injection at runtime; no built-in red-team test harness, OWASP LLM Top 10 / MITRE ATLAS automation |
| Copyright / Proprietary-Code Mirror Detection |
Webhook-DLP integration surface available; no first-party similarity-scoring engine |
| GPU Multitenancy + Quota |
No GPU memory limit per NHI/app (inference-runtime concern) |
| Hardware Telemetry (GPU temp / power / PCIe) |
Not in scope; inference runtime + node-exporter cover this |
| Confidential Computing (TEE / NVIDIA CC / Intel TDX) |
Deploy on TEE infrastructure; not a product feature |
| Model Weight Encryption at Rest (HSM/TPM) |
Inference-platform responsibility |
| Micro-segmentation / East-West SDN |
Kubernetes/SDN layer; AgentWatch consumes mTLS but doesn’t orchestrate the mesh |
| Multi-Region Failover with 99.99% Uptime SLA |
Architecturally supported via stateless app tier + HA Postgres + multi-region LB; the SLA is the customer’s deployment contract, not a vendor commitment from the product |
| Auto-Forward of Test Results to External QA Systems |
Webhooks can forward; no dedicated test-result reporter |
| OpenAPI 3.x Import + Developer Portal Try-It-Out |
OpenAPI generation ships (Swagger UI for gateway routes); OpenAPI 3.x import for cataloguing arbitrary 3rd-party API products is not native |
| Interactive Code-Snippet Generation (Python / Node / LangChain / LangGraph) |
OpenAPI spec lets SDKs auto-generate; no first-party multi-language snippet exporter |
