[object Object]

Why Multi-Vendor

Enterprises run many AI agents — Salesforce Agentforce, Microsoft Copilot, Zendesk AI, ServiceNow Now Assist, plus custom builds on Anthropic, OpenAI, and Google models. They need to coordinate without one vendor owning the orchestration. Agent Fabric is Salesforce’s answer: a neutral control plane that treats all agents as first-class citizens. The TDX 2026 announcement extended Fabric beyond Salesforce-native agents to any agent that registers via MCP or A2A (Agent-to-Agent) protocol, which means you can route a Copilot task and an Agentforce task in the same workflow without writing glue code.

Discovery

Expanded agent and MCP server discovery. One catalog of what agents and tools exist across your stack, with metadata on capabilities, cost per invocation, latency p95, and the data classifications each agent is cleared to read. No more tribal knowledge of “which agent does what.” The catalog publishes JSON-LD descriptors that look like this:

{
  "agentId": "msft.copilot.sales.v3",
  "capabilities": ["account-research", "email-draft"],
  "data_classes": ["public", "internal"],
  "cost_per_call_usd": 0.012,
  "p95_latency_ms": 1800
}

Pair the catalog with a search-by-capability API so a workflow author can ask “give me an agent that can summarize a meeting and is approved for confidential data” and get a ranked list.

Deterministic Orchestration

Critical for enterprise. Coordinate agents with guaranteed ordering and compensating actions on failure. Not every workflow benefits from “LLM decides” — some need deterministic guarantees on sequence, idempotency, and rollback. Agent Fabric supports both modes: free-form (a planner LLM picks next steps) and scripted (a Flow-style DAG with explicit transitions). The scripted mode compiles down to a state machine with at-least-once delivery, dead-letter queues, and Saga-pattern compensations — the same primitives you would expect from Temporal or AWS Step Functions, but with native CRM context.

LLM Governance

Policy at the Fabric layer — which LLMs which agents can use, rate limits, cost budgets, data-access boundaries. Central controls regardless of which agent framework spun the task. A typical policy:

policy: finance-team-agents
  allowed_models: [claude-sonnet-4.5, gpt-4.1-enterprise]
  denied_models:  [any model not in BAA-covered list]
  budget_usd_per_day: 250
  data_classes_allowed: [public, internal, confidential]
  pii_redaction: required

Violations are denied at the Fabric proxy, logged to the Trust Layer, and surfaced in the Agentforce Command Center dashboard.

When to Use Fabric vs. Direct Calls

Use Fabric when more than one team owns agents, when you need cross-vendor routing, or when audit and budget enforcement must be central. Skip it for a single-team, single-vendor pilot — the overhead is real and the policy surface is unused.

What to Do This Week

Inventory the agents your org already runs (Salesforce, Microsoft, custom) and register the top three in the Agent Fabric catalog with cost and data-class metadata.

[object Object]
Share