Harness Overview | Moveworks

The harness is everything around the language model: the runtime that assembles what the model reads on each turn, runs the reasoning loop, calls tools, keeps context and memory, applies safeguards, and produces the reply an employee sees. The model predicts the next token; the harness turns that prediction into finished work. This area explains how that machinery works and where its behavior can be shaped.

Read it when you want to know why the assistant answered the way it did, when you need to steer a behavior, or when you are deciding how to build on the platform.

The Reasoning Engine and Its Upgrade

The Agentic Reasoning Engine powers the main Moveworks AI Assistant. It is the engine behind plugin selection, memory constructs, guardrails, and the model upgrade tracks that each tenant controls. Most pages in this area describe it.

The Agentic Orchestrator is the upgrade to the Reasoning Engine; Productivity Assistants get it first.

Controlled Availability

The Agentic Orchestrator is in Controlled Availability and comes to your Enterprise Assistant in Q1 2027. Capabilities, limits, and interface labels described here may change before general availability. To request access, contact your Moveworks account team.

Start Here

The Reasoning Engine and the Orchestrator are built around the same constraints: a model that predicts the next token, a context window with a hard limit, and attention that fades in the middle of a long input. The engine pages assume you know this, so read LLM Fundamentals first if you are new to language models.

What Goes Into a Turn

What the reasoning engine reads on each turn, the context it draws on, and a map of the reasoning loop.

Agentic Orchestrator

The upgrade to the Reasoning Engine: what it adds, how a turn runs, and where to start.

LLM Fundamentals

Next-token prediction, context windows, and the attention curve that shape every design decision on the platform.

Routes by Job

Understand a behavior

When a response from the main AI Assistant surprises you, check two things first: the AI Reasoning for the response (open More info > Reveal AI Reasoning on it), which shows whether a plugin was called, and the citations panel. A response with no plugin call and no citations was likely ungrounded. The pages below explain the mechanisms behind what you see there, along with Conversation History for how prior turns influence a response and Processing Tool Responses for what the engine does with large plugin outputs.

Grounding and Hallucinations

Why a response may lack citations, how to trace its sources, and what Moveworks does to keep answers grounded.

What Goes Into a Turn

Org knowledge, available tools, this conversation, and results so far: the four kinds of context the engine reads.

Tool Selection

How the engine chooses tools from descriptions and examples, and when it calls more than one.

Steer the assistant

Behavior in the main AI Assistant is shaped from three places. Plugin-level instructions live with the plugin in Tools; see Steering Conversational Plugins. Disclaimers, labels, and toxicity filter instructions are set in Moveworks Setup; see Configure Disclaimers & Labels. Organization-wide steerability instructions are experimental and are applied by Moveworks Implementation and Support teams.

Steerability Tools

Global and situational instructions that adjust tone, terminology, plugin priority, and resource preference.

AI Disclaimers

The three optional disclaimer messages, where each appears, and how they set expectations for users.

Guardrails

Input and output safeguards, including the toxicity filter, sensitive topics policy, and citations.

Build on the Orchestrator

An assistant on the Orchestrator is configured in layers. Instructions are read in full on every turn, Skills load on demand when a request matches their description, and tools supply the data and actions. Built-in code execution and artifacts come with every assistant. If your team already has skill bundles written for another agent host, see Convert an Existing Skill to the Agentic Orchestrator.

Skills

Packaged procedures that load on demand, and the Instructions that route to them.

Planning

How the Orchestrator answers directly, calls tools, or drafts a plan the user approves.

Tool Orchestration

How Tool Studio plugins, MCP servers, A2A agents, and built-in tools reach the model, and how results are processed.

Choose an upgrade track

New reasoning engine model versions reach the main AI Assistant on one of three tracks: Frontier, Standard, or Basic. The choice applies to a whole tenant and can differ between your tenants, so a sandbox can run ahead of production. The companion pages explain which models do which work and how the assistant improves between releases.

Reasoning Engine Upgrade Program

The three upgrade tracks, where to select one, and the behavior changelog for each model version.

Models

How foundation and task-specific models divide the work, how Moveworks improves them, and why nothing you say in chat retrains them.

Where Harness Fits

The harness decides; tools act. Everything either engine calls, including plugins, MCP servers, A2A agents, and connectors, is built and maintained in Tools, and Plugins are where deterministic logic belongs when a step must run the same way every time. Employees meet the result in the Assistant Platform, in the Enterprise Assistant and in Productivity Assistants.