Salesforce engineers share how they built Agentforce, an enterprise AI agent runtime that balances LLM flexibility with deterministic reliability. The core challenge: LLMs interpret language plausibly but enterprise workflows demand enforceable guarantees. Their solution, called 'guided determinism,' uses Agent Graph orchestration to model business processes as graphs with hard validation gates, preventing models from reasoning past critical checkpoints like identity verification. Specialized subagents handle distinct tasks (routing, verification, escalation) with different latency and trust profiles, while smaller fine-tuned models (8B–32B parameters) handle routing at ~50ms latency instead of expensive frontier models. The architecture introduces the concept of 'subjectiveness budgets' — allocating model freedom only where it adds value and enforcing determinism where correctness is non-negotiable. Observability tooling lets engineers trace exactly which subagent took which action in production.
