Architecting Agentforce

Most Agentforce cost overruns are not a pricing problem. They are an architecture problem.

This field manual is a practical guide to designing AI agents that spend Flex Credits deliberately. The core idea is deterministic-first: use the Salesforce platform's existing tooling, Apex, Flows, integration, and deflection, for everything that has a known answer, and reserve generative reasoning for the genuinely ambiguous work where it earns its cost. The result is agents that are cheaper to run, easier to test, and more predictable in production, without giving up the capability that makes them worth building.

Written from the field by a working Salesforce architect, it covers the full lifecycle:

Inside: The Agentforce cost model, and why action count usually matters more than prompt length Deflection patterns that resolve common requests before the planner ever runs Sub-agent and orchestration design that keeps reasoning scoped and observable Compliance, security, and data governance for regulated environments A testing pyramid and CI/CD release gates that catch cost and behavior regressions early Operational guardrails, cost telemetry, and a go-live checklist Worked end-to-end examples that contrast a naive build with an optimized one

It is opinionated on purpose. Treat it as a set of defaults to argue with, grounded in real enterprise delivery, rather than official product documentation.

For Salesforce architects, platform engineers, and AI program leaders who own Agentforce outcomes and answer for what they cost.