Shivanath Devinarayanan · Updated
On this page
Agentforce is the Salesforce runtime. Architecture, identity, network, governance, and observability have to sit around it, or the org gets a demo that cannot be reviewed. Inventory, named owners, a governed gateway, and traces are how a fleet stays in production.
I run this operating model across 200+ production agents. That is the allowed operating fact. These pages do not publish employer return figures, talk-only dollar impact, or a live fleet count. Job title lives on Home and About. This is method, not a customer story.
The nine services stay the paid catalog. This hub is how the work is thought, not a tenth SKU. Hiring still starts at stand up production AI agents , then contact .
The governed path a request walks.
Registry, orchestration, tools, and surfaces.
Policies, owners, gateway, sensitive work.
Traces, cost per run, and remaining room.
The operating model
Architecture
Seven layers, one governed path. Interaction, runtime, orchestration, tools, gateway, memory, and systems of record. Tool and model calls share one front door.
Setup
Enable Einstein and Agentforce, stand up one Connected App, create the first Service Agent, and test in Agent Studio before anything else calls it.
Identity
On-behalf-of is the difference between a shared robot user and a reviewable actor. Tokens narrow hop by hop. The gateway does OBO, not the wrapper.
Data
Data 360 is the system of context. Agentforce is the system of agency. Govern the store before the agent, or production replays whatever the store got wrong.
Documents
Classify, extract, validate, integrate, then a named reviewer. A file is not a tool until a person can audit the fields it produced.
Network
The fleet: Agentforce agents, brokers, MCP servers, and the surfaces that call them, registered and routed as one. One gateway, one registry, copyable policy.
Governance
Policy families, named owners, a governed gateway, and isolation for restricted work. This is the method. The paid offer stays on the services catalog.
Observability
Traces, cost per run, live capacity, and attribution. Exporting logs and building a private dashboard is not observability. If you cannot replay the run, it is not production.
Topics
Architecture, setup, identity, data, and documents sit next to this hub. Network, governance, and observability nest one level down. Every topic stays linked from this article so a crawler that never opens the nav still walks the operating model.
- Seven layers, one governed path
- Agentforce setup in the org
- The agent acts as a named person
- Context before agency
- Files become tools only after review
- Agentforce Network
- Name every agent before you scale
- Guided determinism
- Agentforce as a peer
- MCP servers sit behind the gateway
- Meet the company where it already talks
- Governance is how it stays in production
- Four policy families
- A person owns the decision
- One front door
- Isolation for restricted records
- If you cannot replay the run, it is not production
- Traces and decision paths
- Cost is per agent run
- Remaining room, not a post-mortem
How to hire
Hiring starts with a governed production agent on one documented job, then a consultation.
Start at stand up production AI agents , then book a consultation . These pages are the operating model. The nine services remain the paid catalog.
SHIVANATH DEVINARAYANAN · SALESFORCE MVP HALL OF FAME ·
SD
SHIVANATH DEVINARAYANAN · SALESFORCE MVP HALL OF FAME ·
SD
Stand up the first production agent
Hiring starts with a governed production agent on one documented job, then a consultation.
Also in this file: Home · About · Services · Portfolio · Success Stories · Writing · Events · Contact · Media kit · Privacy Policy