PLATFORM LAYER · SENTINEL-WIDE
Sentinel Progressive Discovery Spine.
The architectural layer that lets enterprise AI work against ERP data at scale — without context blowout, without hallucinated tool selection, without audit gaps.
The reason Sentinel scales from customer 1 to customer 300 when competitors stall at pilot.
"Across dozens of integrations, a significant portion of the context window is consumed before the model does any actual reasoning. Teams blame MCP when the issue is implementation."
— David Soria Parra, co-creator of MCP · Anthropic · MCP Dev Summit NA 2026
The Problem · and the Pattern
Raw MCP at enterprise scale
What breaks without PDS
- 200+ ERP tools dumped into agent context
- 20%+ context window gone before reasoning starts
- Models hallucinate wrong tool choices
- No retry, observability, or backpressure
- Cross-tenant data leakage risks at scale
- Zero audit trail for SOC 2 compliance
- POC works. Customer 3 fails.
Sentinel Progressive Discovery Spine
What PDS delivers
- One entry point: search_tools(nl_query)
- 5–8 tools loaded per task, not 200
- Semantic entity tools, not raw ERP tables
- Retry + backpressure + observability built in
- Tenant isolation enforced at protocol layer
- 100% audit log · SOC 2 Type II ready
- Customer 1 and customer 300 run the same way
Architecture · where PDS sits
SENTINEL PRODUCTS Canopy · Cascades · Foresight · Command Center · Hawk · Warden
↓ (semantic tool invocations)
PDS search_tools · package loader · session cache · gateway · audit
↓ (CDM-normalized queries)
CONNECTORS Summit · Ridgeline · Timber · Elder · NetSuite · Dynamics · Mfg · CSV
↓
CUSTOMER ERP Oracle · SAP · Epicor · AS/400 · D365 · NetSuite · MES · flat files
Ten architectural principles
01
Semantic entity tools
search_open_pos, get_supplier_scorecard — composed business queries, not raw tables.
02
Workflow-scoped packages
Canopy / Cascades / Foresight packs — 5–8 tools per task. Never 200.
03
Tool search as default
One meta-tool stays loaded. Agent asks in plain English, PDS retrieves the right five.
04
Normalized data model
CDM abstracts Oracle vs SAP vs Epicor. Agent reasons on entities, not ERP schemas.
05
Gateway control layer
Retry, backpressure, circuit breaker, tenant isolation, schema validation — built in.
06
Per-session caching
SAP calls take seconds. Cache per session with freshness metadata — agents decide when to bypass.
07
Tenant-scoped catalogs
Only tools for the active customer's ERP are exposed. Smaller context, zero collisions.
08
Failure-aware descriptions
SLA metadata on every tool: latency, freshness, batch windows. Agents plan intelligently.
09
Action menu = curated invocation
Right-click any entity → 6 pre-parameterized workflows. Zero hallucination surface.
10
Ask-Sentinel as front door
One NL query bar replaces 200+ tool selection. The spine resolves the rest.
SLAs · what PDS commits to
< 5%
context used by tool defs
60%+
cache hit per session