The Agentic Reasoning Engine

The Agentic Reasoning Engine powers your Enterprise Assistant, the main Moveworks AI Assistant. A language model reasons over the user’s request and calls tools in a loop: it plans the steps toward a goal, selects and runs the next action, looks at the result, and adjusts the plan.

New to LLMs? Start with LLM Fundamentals for how language models handle context windows and attention.

What it is built from

The reasoning loop. The engine plans the most useful course of action, runs the plan step by step, and shows the user its progress, asking for confirmation when it needs to.
Tools. Search, workflows, API calls, code, and summarization, built in or added through Tool Studio.
Context. On each turn the engine reads your organization’s content and terminology, the tools available to the user, the user’s attributes, and the current conversation.
Models. Large foundation models for reasoning and planning, and smaller task-specific models for jobs like language detection, toxicity checks, and translation.
Guardrails and governance. Safeguards that keep interactions safe and accurate, plus visibility into what employees ask for.

Where you shape behavior

Tool descriptions, examples, and reasoning instructions in Tools.
AI disclaimers and toxicity filter instructions in Display Configurations in Moveworks Setup.
Which user attributes the assistant uses to tailor answers, in Personalization.
How long conversation history lasts, in Conversation History.
The model version, through the Frontier, Standard, or Basic track in the Reasoning Engine Upgrade Program.

The Agentic Orchestrator is the upgrade to the Reasoning Engine and comes to your Enterprise Assistant in Q1 2027.

How it works

What Goes Into a Turn

What the engine reads before each step, and a map of the loop with links to each stage.

The Reasoning Loop

How the engine plans, runs, and checks its work, including Tool Selection and Multi-Tool Responses.

Processing Tool Responses

What happens to a tool’s output before the engine reasons over it, including size limits and variable tracking.

Models

Which models the assistant uses, how Moveworks evaluates and improves them, and why chatting with the assistant does not retrain them.

The Golden Rule

Why attention limits matter for your tools, and how to move logic out of the LLM into code.

Multilingual Support

How the assistant detects and responds in the user’s language.

Context management

Conversation History

What carries over between messages, how long it lasts, and how to clear it.

Personalization

How user attributes such as location and department tailor answers from Enterprise Search.

Domains

How the engine understands your organization’s areas.

Entity Catalog

How the engine understands your organization’s terminology.

Guardrails

Guardrails covers the toxicity filter and conversational safeguards. Related pages:

Built-in skills

Reasoning Engine Upgrade Program: the three upgrade tracks and the behavior changelog for each model version.