AI Executive Office — CXO Assistant Platform  ·  View 16 of 30  ·  5 · Runtime

Agent Orchestration

Who does what across a turn — and why the specialist agents are capability packs rather than separate applications.

Editable source SVG draw.io All views
Interpret Plan Ground Compute Compose Verify Executive Orchestrator Classify intent small model Resolve entities against canon Select capabilities Set budget tokens · tools · time Merge agent results Answer or abstain Specialist agents Finance agent Projects agent Choose tools from its own catalogue Interpret results Domain narrative Tool plane Typed contract no free-form SQL Scope check caller, not service Execute read Mint evidence id Replay for audit Data and models Entity lookup graph Search index Semantic model ML endpoint forecast · anomaly Graph traversal Azure OpenAI reasoning model Guardrails Prompt shield Tool allow-list per role Sensitivity filter Claim binding Groundedness gate Trace + cost record Agent Orchestration — Who Does What in a Turn Agents are capability packs inside one runtime, not separate applications. Adding a domain adds a tool catalogue and a prompt pack, not a deployment. v 1.0 · owner Data & AI Global Practice · date 2026-09

The decision

  • Specialist agents are prompt packs plus tool catalogues inside one runtime. Adding a domain adds configuration, not a deployment, which is what makes an agent marketplace a product feature rather than a programme
  • The orchestrator sets a budget — tokens, tool calls and wall-clock — before any agent runs. Without it, a multi-agent system's cost per question is unbounded and unpredictable, which is fatal to a per-tenant commercial model
  • Guardrails run in every column, not as a final check. Prompt shielding at input is worthless if retrieved document content is later treated as instruction

Risks

  • Multi-agent orchestration is the most over-used pattern in enterprise AI. Most executive questions need one agent and two tools; the architecture supports fan-out but the router defaults to the cheapest path (view 17)
  • Agent-to-agent chatter is where latency and cost hide. Agents return structured results to the orchestrator and do not talk to each other

Assumptions

  • Ten capability packs at maturity; three in the prototype. The packs are versioned with the tenant configuration, so a tenant can be behind on a pack without being behind on the platform
◀  Critical Flow — One Question, End to End All views Who Answers What  ▶
Solution Architecture v1.0 · Azure-native, multi-tenant, sovereign-capable · Data & AI Global Practice · 2026-0916-agent-orchestration