COPILOT STUDIO — PROJECT 6🟠 ADVANCED
OrchestrAI-Agent — New Orchestrator's Agentic Reasoning Loop handling a multi-step deal-desk request. Static reference showing the plan → act → observe → re-plan trace and its cost.

🔁 Agentic Reasoning Loop — "Prepare Deal Package for Acme Corp"

PlanBreak request into: check standing → draft summary → check shipping
ActTool: Customer Standing Lookup (MCP) ~5 cr
ObserveResult: Account in good standing, tier: Enterprise
ActTool: Discount Approval Drafting (Prompt Builder) ~14 cr
ObserveDraft generated — requested discount: 25%
Re-plan25% exceeds Skill guardrail (max 15% autonomous) — halt & escalate
ActTool: Shipping Options by Region (Custom Connector) ~5 cr
SynthesizeCombine results into one coherent reply ~6 cr

Skill: "Prepare Deal Package"

Trigger: "prepare a deal / proposal for [customer]"
Tools: Standing Lookup → Discount Drafting → Shipping Lookup
Guardrail: never auto-approve a discount above 15% — always escalate to human sign-off.
🛑 Guardrail triggered this run — 25% discount routed to manager approval instead of auto-completing.

Model configuration

Orchestration typeNew Orchestrator
Deep reasoningOn
Drafting modelBring Your Own Model (Foundry)
~30 Copilot Credits
for this single multi-step turn — vs. ~2 credits for a Project 1-style simple grounded answer
What this shows: the New Orchestrator autonomously chained three tools and a premium reasoning/drafting pass to fulfill one request — and its own Skill guardrail correctly stopped it from auto-approving a discount above policy, routing to a human instead. The cost panel makes the "premium reasoning is expensive" lesson concrete rather than abstract.