Reasoning Plane

Where AI reasons and acts on enterprise data — a ReAct (Reason + Act) agent loop powered by LLMs, with registered tools that can query data, execute workflows, and interact with external systems under policy guardrails.

For platform-overview framing see Platform Architecture > Reasoning Plane. For loop mechanics and safety guarantees see AI Operating System > ReAct Layer.

Capabilities

• AI Copilot — natural-language interface for querying data, running analysis, and executing actions.
• ReAct Tools — registered tool functions the AI can call.
• Planning Playground — RAG validation and testing environment for reasoning chains.
• Policy Guardrails — pre-execution policy checks before any tool action runs.
• Multi-system Orchestration — coordinate actions across multiple systems in a single reasoning chain.
• Inspectable agent traces — per-step tool call, latency, hits, and token cost surfaced for traceability.

ReAct Loop

graph LR
    User["User Query"] --> Copilot["AI Copilot"]
    Copilot --> Reason["LLM Reasoning"]
    Reason --> Plan["Plan Actions"]
    Plan --> Policy["Policy Check"]
    Policy --> Execute["Execute Tool"]
    Execute --> Observe["Observe Result"]
    Observe --> Reason
    Execute --> Systems["Enterprise Systems"]

Registered Tool Categories

• Data Plane queries — query virtual entities across connected systems.
• Ontology queries — search the knowledge graph and semantic index.
• Journey tools — execute and manage data-processing journeys (Order-to-Cash, Demand Forecasting, etc.).
• Logs & Reports tools — query execution logs and generate reports.
• Solution-app tools — per-Solution app-curated tools (e.g. for Order-to-Cash: Order Match Verifier, AR Cash Flow Projector, O2C Alert Generator, Dunning Notice Composer).
• Write-capable tools — high-risk-write actions (e.g. create_salesforce_lead) gated by approval.

Console Walkthrough

Four tabs cover interactive copilot use, RAG validation, the tool registry, and custom tool authoring.

AI Copilot

• Natural-language input box with an expandable agent-steps card per response.
• Per-step trace: tool call name, latency, retrieval hits, relevance scores.
• Read-only / write badge and token-cost surfaced for traceability.

Planning Playground

• Semantic RAG query input with quick-fill scenario chips.
• Tabs for Retrieval Bundle, Graph View, and Test Suite.
• Retrieval Quality Score with per-dimension breakdown so engineers can iterate.

ReAct Tools

• Per-tool card: purpose, read-only or write tag.
• Built-in tools cover schema discovery, query, semantic search, and Solution-specific actions.

Tool Builder

Related References

• ReAct Layer — loop mechanics, policy gating, failure modes.
• Semantic RAG — what the retrieval bundle must contain to enable safe orchestration.
• Implementation Guide — production code (agent_executor.py, policy_engine.py, tool_registry.py, intent_analyzer.py, ...).
• Console > AI Analytics — dedicated console drill-down.
• Control Plane — the semantic layer this plane queries on every cycle.