Product Roadmap (Tier 2 / Tier 3 / Tier 4 + Out-of-Scope)
This appendix captures the forward roadmap. Tier 1 + A2A are shipped in v2.0.45 — see §3.6 (cross-protocol translation), §16 (LDAP group routing), §17 (per-agent tool whitelists), §18 (Presidio bridge), §19 (guardrail versioning), §20 (compliance + framework selector), §22 (cost anomaly events), §28 (HITL workflow events), §31 (backup), §35 (secret manager integrations), §36 (network controls), §43.1 (acknowledged gaps), and the new Tier 1 + A2A files under src/enhanced/tier1/ + src/enhanced/a2a/.
The items below are not shipped and are the roadmap for v2.1+.
50.1 Tier 2 — Strong Differentiators (Quarter 2–3)
Section titled “50.1 Tier 2 — Strong Differentiators (Quarter 2–3)”Closes the gap with specialized AI-observability + agent-platform tooling (Datadog AI, LangSmith, Helicone, Langfuse pro tier).
| Capability | Why it matters | Effort |
|---|---|---|
| Prompt Registry — Dev/Staging/Prod versioning, tagging, rollback, audit trail | Enterprise prompt-management table-stakes; lets teams version-control prompts alongside code | M |
| Prompt Evaluation Framework — offline + online test sets, scoring rubrics | Engineering teams adopting AgentWatch want systematic quality measurement | M–L |
| Prompt Playground with interactive testing + RBAC + sharing | Sales-demo + developer-onboarding accelerator | M |
| A/B testing / canary deployment for prompts + models with traffic splitting | Production traffic experimentation per route | M |
| Virtual MCP servers (curated tool-subset views over base servers) | Fine-grained MCP access governance without forking servers | S–M |
| RAG quality metrics — Context Precision / Recall / Faithfulness per request | Pair with a Ragas integration (webhook-DLP-style hook) | M |
| Hallucination / faithfulness monitoring — LLM-as-a-judge on outputs | Datadog AI Observability has this; AgentWatch parity | M |
| Topology / lineage visualization (Agents → MCP → Data Sources) | Strong sales-demo signal; pairs with the existing Knowledge Graph viz | M |
| OpenAPI 3.x import + Developer Portal “try-it-out” | Generation side already ships; import side missing | M |
| Interactive code-snippet generator (Python / Node / LangChain / LangGraph / cURL) | Drop-in production wiring for validated prompts + tools | S–M |
| Automated kill-switch — policy-driven termination on anomaly | Tier 1 cost-anomaly alerting feeds this; closes the loop | S |
| Model drift detection — refusal-rate baseline + statistical drift alerts | Quality-degradation alarms | M |
| Prompt drift analysis (system-prompt change tracking) | Pairs with §Tier1 config drift monitoring | S–M |
| Persistent agent memory — short-term + long-term cross-session | Pattern from LangGraph / Mem0; LanceDB already in place | M |
| Context-window usage visualization (“lost in the middle”) | Token counts already tracked; just needs the analytical lens | S |
| Acryl DataHub lineage integration | Native lineage feed; Kafka already enables external reconstruction | M |
| LLM-as-a-judge jailbreak / indirect-injection detection | Matches Datadog AI Guard sophistication; complements existing pattern DLP | M |
Sequencing recommendation: Prompt Registry → Prompt Evaluation Framework → Playground → A/B testing → RAG metrics. The Registry is the dependency anchor for everything else.
50.2 Tier 3 — Strategic / Multi-Quarter
Section titled “50.2 Tier 3 — Strategic / Multi-Quarter”These reshape AgentWatch’s architecture. Build only when a major customer commits or when a strategic-bet decision is made.
| Capability | Why it matters | Effort |
|---|---|---|
| Declarative policy engine — OPA / Rego or Google CEL embedded | Replaces ad-hoc RBAC config with versioned, testable policy code | L |
| NHI governance — registration, owner mapping, RFC 8705 cert-bound tokens, review cadence | Non-human-identity is the next enterprise IAM frontier | L |
| Okta-anchored NHI lifecycle | Pairs with NHI governance | L |
| Multi-agent orchestration — recursion-depth limits, agent profile mgmt | Agentic AI is the next deployment pattern after pure LLM gateways | L–Strategic |
| Agent Registry + versioning + safe rollback — agent-as-artifact concept | Depends on multi-agent orchestration | L |
| Framework-level tracing auto-instrumentation — LangGraph, CrewAI, OpenAI Agents SDK, AWS Bedrock Agent SDK | Matches Datadog’s strongest differentiator (§45.3) | L |
| Adversarial / red-team CI framework — OWASP LLM Top 10 + MITRE ATLAS test corpus | AI-security-test parity with Lakera / Robust Intelligence | L |
| A2A v2 — RFC 8693 token exchange (on-behalf-of), push notifications, external-agent pass-through proxy | Extends the §A2A v1 scaffold to the full design spec | L |
| SPIFFE / SPIRE workload identity | Hardware-attestation identity for on-prem clusters | Strategic |
| Permission intersection model (user role ∩ agent role) | Foundational for NHI governance | M–L |
| Prompt compression (token-reduction algorithms before LLM) | Pairs with prompt + semantic caching for compounded savings | M (research-heavy) |
| Unit economics + forecasting | Predictive cost modeling | M–L |
| Copyright / proprietary-code similarity detection | ML similarity engine OR vendor webhook | M |
50.3 Tier 4 — Lower-Priority Roadmap Items
Section titled “50.3 Tier 4 — Lower-Priority Roadmap Items”Documented for completeness; smaller scope, often quality-of-life.
| Capability | Effort |
|---|---|
| Native Terraform provider (replaces Git-version-controlled JSON config) | M |
| White-label SaaS branding for resellable multi-tenant UI | S–M |
| Auto-forward test results to external QA systems (beyond generic webhooks) | S |
| Configuration validation tool against vendor-spec | S |
| Sandbox & Developer Platform (composite of Prompt Playground + snippet gen + tool config sharing) | L |
50.4 Out-of-Scope by Design (Not Gaps — Infrastructure-Layer)
Section titled “50.4 Out-of-Scope by Design (Not Gaps — Infrastructure-Layer)”These appear in customer RFPs but should be answered “deploy on the appropriate platform layer,” not “we’ll build this.” Repeating from §43.2 for completeness in the roadmap context:
| Concern | Owns it |
|---|---|
| GPU multitenancy / VRAM quotas | Inference runtime (vLLM, Triton, TGI) |
| Hardware telemetry (GPU temp, power, PCIe) | node-exporter + IPMI exporters |
| Confidential Computing / TEE | AWS Nitro Enclaves, Azure Confidential Computing, Intel TDX, NVIDIA CC |
| Model weight encryption at rest (HSM/TPM) | Inference platform + cloud KMS |
| Micro-segmentation / east-west SDN | Kubernetes NetworkPolicies, Istio AuthorizationPolicy, AWS Security Groups |
| mTLS service mesh orchestration | Istio, Linkerd, Consul Connect (AgentWatch consumes mTLS, doesn’t orchestrate the mesh) |
| Multi-region 99.99% uptime SLA | Customer deployment contract (AgentWatch is architecturally capable; the SLA itself is operational) |
50.5 Recommended Sequencing
Section titled “50.5 Recommended Sequencing”Q1 (v2.1) — Polish + Tier 2 RAG/Eval anchor:
- Prompt Registry (Tier 2)
- RAG quality metrics integration (Tier 2)
- A/B canary deployment (Tier 2)
- Throttle UI + topology visualization (Tier 2)
Q2 (v2.2) — Prompt + Agent lifecycle:
- Prompt Evaluation Framework
- Prompt Playground + code-snippet generator
- Virtual MCP servers
- Persistent agent memory
- Hallucination + LLM-as-a-judge detection
Q3 (v2.3) — Strategic AI-platform expansion:
- Multi-agent orchestration v1
- Framework-level tracing auto-instrumentation
- A2A v2 (token exchange + push notifications)
Q4 (v3.0) — Identity + governance foundations:
- Declarative policy engine (OPA/CEL)
- NHI governance
- Agent Registry
- Adversarial / red-team CI
Strategic, off-cadence:
- SPIFFE / SPIRE
- Prompt compression
- Unit economics forecasting
By end of Q3, the RFP-compliance rate from §46.8 moves from 46% Full Comply / 11% Intend / 43% No Comply to approximately 75% Full / 15% Intend / 10% No Comply. The remaining 10% are the infrastructure-layer items in §50.4 (truly out of scope by design).