[object Object]

Why Multi-Vendor

Enterprise reality: Salesforce Agentforce handles sales and service. Microsoft Copilot Studio handles employee productivity inside Office. ServiceNow Now Assist handles ITSM. A custom agent handles a domain-specific workflow no platform vendor will build. All need to coordinate or at least not step on each other. Agent Fabric is the control plane that gives a security and operations team one place to discover, govern, and orchestrate them. The TDX 2026 expansion added support for any A2A-compliant agent and any MCP-compliant tool, regardless of vendor.

Discovery

Expanded agent and MCP server discovery. The Fabric registry answers “what agents and tools exist in my stack” with metadata on capabilities, cost per call, latency p95, data classes the agent is cleared to read, and policy boundaries (which models, which regions). No more tribal knowledge of which agent does what or which team owns what. The registry is queryable via REST and via the new Fabric MCP Inspector tool that ships in Slack and Tableau Next as a chatbot front-end.

Registry record (JSON):
{
  "id": "msft.copilot.sales.v3",
  "vendor": "Microsoft",
  "capabilities": ["account-research","email-draft","meeting-prep"],
  "data_classes": ["public","internal"],
  "regions": ["us-east","eu-west"],
  "models_used": ["gpt-5-enterprise"],
  "p95_latency_ms": 1900,
  "cost_per_call_usd": 0.014,
  "owner_team": "Sales Operations",
  "policy_id": "sales-team-default"
}

Deterministic Orchestration

Some workflows demand guaranteed ordering, idempotency, and compensating actions on failure. LLM-driven coordination is probabilistic; Fabric’s deterministic primitives give enterprise-grade guarantees on top of LLM agents. Use deterministic mode for compliance-sensitive flows (anything touching billing, legal, regulated data) and LLM-coordinated mode elsewhere where adaptability outranks predictability. The deterministic mode compiles to a state machine with at-least-once delivery and Saga-pattern compensations.

Deterministic step definition (YAML):
steps:
  - id: enrich
    agent: msft.copilot.research
    on_failure: skip-and-log
  - id: draft-email
    agent: sf.agentforce.sales-email
    requires: [enrich]
    on_failure: human-handoff
  - id: log-activity
    agent: custom.activity-logger
    requires: [draft-email]
    compensation: rollback-activity

LLM Governance

Central policy: which LLMs which agents can use, rate limits per minute and per day, cost budgets per team, data-class boundaries, region pinning for residency. A red-team-discovered data leak on one agent gets the whole vendor throttled through Fabric until remediation lands. Enterprise control surface that finally matches the way IT governs SaaS contracts and cloud spend.

Policy: finance-team-default
  allowed_models: [claude-sonnet-4.5, gpt-5-enterprise, salesforce-xgen]
  denied_models:  [any not under signed BAA/DPA]
  per_minute_calls: 600
  per_day_budget_usd: 250
  data_classes_allowed: [public, internal, confidential]
  pii_redaction: required
  region_pin: eu-west
  audit_destination: data-cloud-trust-trace

When to Stand Up Fabric

Stand up Fabric when more than one team owns agents, when you need cross-vendor routing, or when audit and budget enforcement must be central. Skip for a single-team, single-vendor pilot. The break-even point sits around 5+ production agents or 2+ vendors.

Common Failure Modes

  • Treating Fabric as optional and discovering shadow-agent sprawl six months later.
  • Defining policies that are too tight at launch — every legitimate agent fails and the team disables Fabric.
  • Letting deterministic and LLM-coordinated modes mix without explicit ownership; debugging becomes impossible.

What to Do This Week

Inventory the agents your org runs across vendors, register the top three in Fabric with cost and data-class metadata, and define one budget policy per team that owns an agent.

[object Object]
Share