The headline promise: "build autonomous agents with clicks, not code." The reality: you're handed a metadata-heavy framework that still requires Apex triggers for anything beyond basic record routing, and the "Atlas Reasoning Engine" is essentially a prompt template wrapper around standard LLM calls with zero visibility into the actual chain-of-thought. Partners I spoke with are billing 60-70% of implementation hours to custom Apex workarounds because the declarative surfaces don't expose the controls enterprise customers actually need — things like multi-step approval chains, external API orchestration with retry logic, or deterministic handoff rules between agents.
Pricing is the other landmine. The per-conversation model sounds clean until you model a 5,000-seat service cloud org with 15 agents running concurrent workflows. One partner showed me their forecast: $280K/year in Agentforce consumption fees on top of existing Einstein licenses, with no volume discount tier until you hit 2M conversations annually. That's not a partner-friendly motion — it's a tax on adoption.
And the certification path? Three exams, $600 each, valid for 12 months. The study materials reference features that aren't GA yet. Two of the three firms I talked to have already d their practice builds because they can't justify the enablement spend when the roadmap keeps shifting — "Spring '25" features keep sliding to "Summer '25" with no comms.
Data Cloud integration is the one area where the demo matches reality, but only if the customer already has Data Cloud provisioned and unified. Which most don't. So you're selling a two-product motion before the agent even runs its first turn.
Bottom line: if you're a partner betting Q1 revenue on Agentforce implementations, have a Plan B. The platform might get there by Dreamforce '25, but right now it's a v1 dressed up as a platform play.
Next Nvidia's latest demo proves the inference stack matters more →
a practical ChatGPT prompt guide, with plenty of directly applicable cases.