Atlas reasoning engine request handling
Key Takeaway
Each user message enters at the subagent declared with start_agent, which is the Agent Router by default. The agent resolves that subagent's instructions from top to bottom with no LLM. If a transition runs during resolution, the agent discards what it already resolved, jumps to the new subagent at once, and builds the prompt from the new subagent's instructions.
The Atlas reasoning engine is the graph-based engine that powers Agentforce agents. Atlas handles each user message in two phases. First it runs fixed, rule-based steps: it moves to a starting subagent and resolves that subagent's instructions in order. Then it sends a prompt to a large language model (LLM), which reasons, runs actions, and writes the response. Knowing where each phase starts and stops explains how routing, transitions, and business rules behave.
What the Atlas reasoning engine is
Atlas is graph-based. You can think of it as a flowchart with nodes, variables, and transitions, so agents follow specific, predictable paths. Like the conductor of an orchestra, Atlas directs how subagents and actions handle a user's request.
Unlike strictly prompt-based reasoning engines, Atlas separates an agent's big-picture workflow from its conversational skills. Atlas uses Agent Script, the language for building agents in Agentforce Builder, to combine programmatic expressions with natural language instructions. The result is called hybrid reasoning. Hybrid reasoning gives you the predictability and control that enterprises need, plus the flexibility and creativity of LLMs.
Beginning in April 2026, agent topics are called subagents. Functionality did not change, so you may see both terms during the transition.