Vrida — AI Capability Gaps

What the market is building for autonomous/agentic ERP/POS/SaaS that Vrida's AI Capability Plane does not yet cover.

Date: 2026-07-06 Source: competitive research across AI-native ERP/POS startups, incumbents adding agents (SAP Joule, NetSuite Next, Dynamics/Copilot, Agentforce, Intuit, Shopify, Square, Lightspeed, Odoo), and horizontal agentic standards (MCP, A2A, ACP/AP2, OWASP Agentic Top 10, OpenTelemetry). Benchmark: Vrida's AI Capability Plane — Part A (14 infra controls), Part B (19 capabilities), Part C (9 governing rules), Part D (15-question module-walk).

Headline

Vrida is ahead of most of the market on internal governance — budget gate (A3), execution ledger (A6), value meter (A7), rollback/compensating-transaction (A12), AI segregation-of-duties (A13), and the C5/C7/C8 governing rules. Few competitors articulate SoD-for-agents or reversible-financial-action contracts this explicitly.

The gaps are almost entirely outward-facing — interoperability, agentic commerce, and a few execution/data-substrate concepts. The stated risk is inward over-engineering while the race is on the outward interoperability/commerce surface.

Only two gaps genuinely touch the schema layer (G5 agent memory, G3 multi-agent handoff). The rest belong to the AI Capability Plane as documentation, or to the API / testing / runtime layers not yet built — and should not be forced into the schema-design runbook.

The 11 gaps

Priority 1 — Outward interoperability & agentic commerce

G1 — Agent-readable catalog + agentic checkout

No component exposes a tenant's catalog/inventory to external buying agents (ChatGPT Instant Checkout / ACP, Google AP2, Shopify Storefront MCP, Square agentic ordering). For a nursery ERP/POS, this is where product discovery and a growing share of orders will originate.

Plane coverage: Not covered.
Layer: API layer (mostly). Minor schema — an "agent-readable / exposed" flag on catalog tables.
Relevance: Highest revenue relevance of all gaps.

G2 — MCP server endpoints

A1 defines a service-only AI access boundary (internal — good), but the market standard is to also expose a governed MCP server so agents (internal and third-party) discover and use tools via a unified, OAuth-2.1-secured interface. NetSuite, Shopify, and Microsoft all ship MCP.

Plane coverage: A1 is inward-only; MCP-as-tool-contract not covered.
Layer: API layer.
Stack note: Implementable on the Node backend with the MCP TypeScript SDK; gate behind A1's service boundary + A5's agent passport.

G3 — A2A / multi-agent handoff contract

A4 (authority ladder) + A5 (agent passport) govern a single agent's authority. Nothing describes agent-to-agent delegation, handoff, or A2A Agent Cards. The market has moved to multi-agent "concerts" (SAP collaborative agents, NetSuite's agent concert, Dynamics multi-agent orchestration).

Plane coverage: Single-agent only.
Layer: Plane + schema (handoff/delegation records, delegation provenance).
Risk mapping: OWASP "insecure inter-agent communication."

Priority 2 — Data substrate for reliable autonomy

G4 — Governed semantic layer / business ontology

"Ask your data" (B16/B17) and reporting (B18) are only as reliable as a semantic layer beneath them. SAP (Knowledge Graph), NetSuite, and Atlan treat this as core. B3 (self-maintaining master data) is adjacent, but a queryable, provenance-citable semantic layer is distinct — and it directly reduces hallucination risk.

Plane coverage: Partial (B3 adjacent, no semantic layer).
Layer: Query / API layer. Minor schema — semantic annotations on tables.

G5 — Agent memory architecture (distinct from RAG) — BUILT

B11 (tenant operating memory) hinted at it; the 2026 pattern is explicit short-term working memory + long-term episodic/semantic memory with a write path and temporal reasoning that pure vector similarity lacks. Memory-poisoning is a named attack surface not currently in A11.

Plane coverage: Schema built — ai.agent_memory (part of the ai module's 2026-07-06 schema-lock, PROJECT_DECISIONS #25). Partial-unique (tenant_id, category, key) WHERE status='active' matches identity.agent_duty_grant's precedent, so a disabled memory can be superseded by a fresh active one under the same key without destroying history. decision_provenance.memory_refs on crm/inventory tables now resolves to a real ai.agent_memory.id (see CROSS_MODULE_CONTRACTS.md § AI). Memory-poisoning defense (A11) and the short-term/long-term working-memory behavior (retrieval, write-path policy, eviction) are still open — this entry covers the schema substrate, not the runtime/plane behavior on top of it.
Layer: Schema (built) + Plane (still open). This was the one gap with real schema-design implications; the schema half is now done.
Stack note: Mem0 or a Postgres + pgvector store fits the Node/TS stack. ai.agent_memory is the plain-Postgres substrate this pass built; a vector-similarity layer on top remains a future add.

Priority 3 — Lifecycle / quality mechanics (market-standard)

G6 — Agent simulation / sandbox (pre-production)

Shopify SimGym, Salesforce simulation, synthetic-scenario testing before deploy. A9 (eval/replay) is post-hoc; simulating against synthetic nursery scenarios (spring rush, frost event, perishable spoilage) before deploying is missing.

Plane coverage: A9 is replay, not pre-production simulation.
Layer: Testing layer — not schema.

G7 — Trajectory evals + CI-gated regression + OTel conventions

The market eval model is three layers (unit evals + LLM-as-judge regression + production trace sampling) plus trajectory/step-level evals (not just final answer), on OpenTelemetry GenAI semantic conventions.

Plane coverage: A9 aligned but should add trajectory evals + CI-gating; A6/A7 should adopt OTel conventions so they interoperate with standard tooling instead of staying proprietary.
Layer: Testing / observability — not schema.
Stack note: Langfuse (self-hostable, OTel-native) fits a cost-conscious multi-tenant SaaS.

G8 — Proactive / scheduled agents as a runtime primitive

Trigger-based and scheduled background agents (Copilot Studio, Sidekick Pulse proactive alerts, Intuit always-on). Capture/route + ambient analyst partly cover this, but "scheduled/proactive agent" as a first-class runtime concept with its own budget (A3) and authority (A4) is underspecified.

Plane coverage: Partial.
Layer: Runtime. Minor schema — a schedule/trigger table.

Priority 4 — Emerging, lower immediate priority for SMB nurseries

G9 — Agent marketplace / third-party agent extensibility

SAP, Microsoft Agent Store, Copilot Studio. Lower priority for a single-vertical SMB product; relevant long-term.

Plane coverage: Not covered — appropriately, for now.
Layer: Future / platform.

G10 — Voice / vision capture + real-time streaming agents

Voice/vision capture (photograph a packing slip → draft PO/receiving, like Lightspeed OCR and Intuit) is a concrete near-term win; general computer-use agents are lower priority.

Plane coverage: Part D D15 asks capture modality (scan/photo/voice/form) — partially covered at the design-question level, not as a built capability.
Layer: Capture / API.

G11 — Standardized agent identity protocols

W3C DIDs/VCs, IETF agent-auth drafts, SCIM-for-agents. A5's agent passport is a proprietary analog; aligning to emerging standards is future-proofing, not urgent.

Plane coverage: A5 covers the concept proprietarily.
Layer: Identity / API.

Gap → layer summary

Gap	Belongs to	Schema-runbook relevance
G5 agent memory	Schema (built) + Plane (open)	Built — `ai.agent_memory`, 2026-07-06
G3 multi-agent handoff	Schema + Plane	Yes — moderate
G1 agentic catalog/checkout	API (mostly)	Minor (agent-readable flag)
G2 MCP server endpoints	API	None
G4 semantic layer	Query / API	Minor (annotations)
G6 simulation/sandbox	Testing	None
G7 trajectory evals / OTel	Testing / observability	None
G8 proactive/scheduled agents	Runtime	Minor (schedule table)
G9 agent marketplace	Future / platform	None
G10 voice/vision capture	Capture / API	Part D D15 partial
G11 agent identity standards	Identity / API	None (A5 proprietary)

Recommended actions

AI Capability Plane updates (documentation — the heavier set)

Formalize B11 into a proper agent-memory subsystem — short-term working memory + long-term episodic/semantic memory, with an explicit write path and temporal reasoning.
Add memory-poisoning to A11 — distinct from prompt-injection.
Add a multi-agent handoff / A2A concept — extend A4/A5 for agent-to-agent delegation (currently single-agent).
Note MCP server exposure as an infra concept — the outward interoperability the market treats as standard (pair with A1's inward boundary).
Note OpenTelemetry GenAI conventions for A6/A7 — so the ledger and value meter interoperate with standard tooling.

Schema-design runbook updates (surgical — keep it lean)

Add an agent-memory question to Section 2.2's schema-translation half — "does this module read from or write to agent/tenant memory, and what persistent memory state does the schema need?" (currently absent).
Extend D11's trigger to include memory-poisoning as an adversarial surface (alongside prompt-injection).

Explicitly NOT added to the schema runbook (avoid over-engineering)

Simulation (G6), trajectory evals / OTel (G7), semantic layer (G4), proactive-agent runtime (G8), MCP endpoints (G2), agentic commerce (G1) — these are API / testing / runtime concerns that do not shape tables. They belong in the AI Capability Plane and in the future API-layer and testing-layer runbooks, not the schema-design runbook.

Stack-fit tooling (Flutter / JS / Node + AI agents)

Concern	Tool	Why
Agent framework	Mastra	The de-facto TypeScript choice; workflow + human-in-the-loop primitives, MCP support, clean Next.js integration
Durable multi-agent graphs	LangGraph.js	Checkpointing, time-travel, HITL for long-running agents (G3)
Observability / eval	Langfuse	Open-source, self-hostable, OTel-native — fits multi-tenant cost model (G7)
Agent memory	Mem0	Token-efficient, multi-signal retrieval; or Postgres + pgvector (G5)
Tool/data interface	MCP TypeScript SDK	Expose governed tools to internal + third-party agents (G2)

Caveats

Protocol landscape (MCP, A2A, ACP/AP2, UCP) is fast-moving and partially overlapping; favor an abstraction layer over hard-committing to one commerce protocol.
Vendor capability claims and ROI figures in the source research are directional/vendor-reported, not independently verified.
Nursery-specific nuance: live-plant inventory is perishable and seasonal, so retail-vertical autonomy (perishable replenishment, spoilage/markdown agents) matters more than generic ERP-finance autonomy.

Last modified: Jul 6, 2026, 7:48 PM PT