Assistant agentic updates

Assistant introduces a more powerful agentic engine and a streamlined user experience. These agentic capabilities in Assistant guide users when completing complex work. Users can choose between fast responses or extended thinking depending on the depth and complexity of a task. When executing, Assistant adaptively plans and iterates using the full context of your enterprise to take on more and refine its work. It will automatically consolidate insights across all knowledge sources and provide responses that are tailored to you based on who you are, what you work on, and who you work with, all using Glean's Personal Graph.

This article summarizes the upgraded Assistant and explains how to manage the rollout within your company. As an Admin, you can control how to manage this rollout, enabling this for your company or a select group of users before October 30, 2025.

This agentic engine runs on GPT-5 or Claude Sonnet 4.6. If you're using Glean Key, you will be automatically migrated to GPT-5. Customer Key customers must manually migrate their models to GPT-5 or Claude Sonnet 4.6 to see a performance difference.

Glean recommends reviewing all of the features in this document to understand the upgraded Assistant changes that go into effect as part of this managed rollout, as well as how to manage the rollout of the new Assistant experience within your company.

Feature Updates

The agentic engine

The agentic engine powers Assistant, adding adaptive planning and full‑context awareness to help get more done.

With adaptive planning, the Assistant plans step by step and updates the plan as it learns—keeping ideas that hold up and dropping those that don't. This makes it a strong fit for creative and strategic work, tasks that need nuance and exploration, and even simply finding the right thing when users aren't sure how to ask.

It may think a bit longer, but it uses your enterprise context to handle complex, multi‑step tasks more effectively. It also selects the right tools for each job to complete work end‑to‑end. And the experience is more interactive, so people can ask follow‑up questions and dig deeper as they go.

Fast and Thinking modes

Assistant now has Fast and Thinking modes, giving users the ability to choose whether Assistant should try to answer the user's question quickly or spend additional time thinking through the question to thoroughly answer it.

note

Thinking mode is the default mode. If you change modes, your selection is maintained.

info

Fast mode uses a smaller set of tools optimized for speed, which may result in fewer citations for some queries. Actions from connected apps (MCP tools) are not available in Fast mode. If a query requires those actions, switch to Thinking mode, or let Adaptive mode fall back to Thinking mode automatically.

Thinking mode performs deeper reasoning and more reliably searches company knowledge, providing more thorough results with citations. It also supports the full set of tools, including actions from connected apps.

Admins can guide Assistant behavior by adding custom instructions in Admin Console → Platform → Assistant → Settings. For example, adding an instruction like "For every question, first check the company's internal knowledge and answer based on it" can help prioritize company sources.

For more details on citation behavior, see Citations.

All Knowledge mode

All Knowledge mode in Assistant combines your company's internal knowledge, web content, and the LLM's built-in knowledge. Assistant automatically picks the most relevant sources for each question, so you don't need to select "Company" or "World" modes yourself.
You can use all sources or filter to specific sources at the start of the query, depending on the task.

Removal of knowledge mode preferences

To simplify the experience and power every query with the latest agentic engine, Assistant now defaults to All Knowledge mode. All searches and chats automatically blend company and world knowledge as needed for each question. The previous dropdown for selecting Company, World, or All has been removed, and users no longer set a per-user mode preference. Users will no longer see these options:

• The selection option for Company, World

• The preference to set their knowledge mode

For control at query time, you can use filters to toggle Search the web and Use company sources to restrict results to external or internal sources, respectively. Use the following selections for equivalent functionality:

Previous setting	Updated setting
Company knowledge
World knowledge
New Pre-trained LLM knowledge only note Users who don't need access to real-time world or company data can now get quick answers directly from pre-trained LLM knowledge.

note

• When both web and company toggles are unselected, Assistant relies on the LLM's pre-trained knowledge to answer questions. Responses generated this way won't include citations because no sources are retrieved.
• Assistant selects the web search tool appropriately when web search is needed; you cannot force a web search.

Enable Agentic Model for the agentic engine

In order to enable the agentic engine, you must use OpenAI GPT‑5 or Claude Sonnet 4.6. GPT-5 is available via Glean Key or Customer Key through OpenAI Direct or through Azure. Claude Sonnet 4.6 is available via Customer Key through Google Vertex AI or Amazon Bedrock.

1. Navigate to Admin Console → Platform → LLMs.

2. Under the Select models for assistant step, pick your Provider, specify GPT-5 or Claude Sonnet 4.6 for your agentic reasoning model. For e.g:

   

note

The new agentic engine is only available in Assistant today. Agents will continue to use the agentic and fast agentic models.

Enable new agentic engine features

You can optionally enable the new simplified chat experience, all knowledge mode, and agentic engine from the Admin Console. You can enable these features for admins only, or roll them out to the entire organization.

caution

Prerequisites

• If you're self-hosted, ensure your deployment has occurred on or after September 4th, 2025
• If you're using your own LLM key, GPT-5 must be enabled

1. Navigate to Admin Console → Platform → Assistant → Settings

2. Under New agentic engine features, enable Fast and Thinking modes

   

note

On October 30th, 2025, these features will be enabled by default for everyone and this setting will be removed.

Chat API integrations

If you use the Chat API to integrate Glean Assistant into your own applications, the agentic engine is not enabled by default for API requests. When you omit agentConfig from a Chat API request, the API defaults to legacy Chat V2 behavior, even if the agentic engine is enabled for your tenant in the Glean UI.

To use the agentic engine through the Chat API, set agentConfig.agent in your request body to one of the following values:

• FAST: Uses the agentic engine with a fast model set and reduced reasoning effort. Corresponds to fast mode in the Glean UI.
• ADVANCED: Uses full agentic reasoning with the best model set. Corresponds to thinking mode in the Glean UI.

If you don't set agentConfig.agent, or set it to DEFAULT, the API uses legacy Chat V2 without the agentic engine.

The same model requirements apply to API integrations: GPT-5 or Claude Sonnet 4.5 must be configured as the agentic reasoning model. Without a supported model, the API falls back to legacy Chat V2 regardless of the agentConfig setting.

Verify the agentic engine is active

To confirm that a Chat API response used the agentic engine, check for "workflowId": "INTERNAL_LOOPING_PYAGENT" in the response body.

Citations in Chat API responses

Citation behavior differs between legacy Chat V2 and the agentic engine:

• Legacy Chat V2 (DEFAULT): May return fewer or missing citations compared to the Glean UI.
• Agentic engine (FAST or ADVANCED): Provides grounded answers with consistent citations, including deep-linked citations when supported.

If your Chat API integration previously returned citations and they stopped appearing after enabling the agentic engine in the UI, verify that your API requests include agentConfig.agent set to FAST or ADVANCED.

For more details on the Chat API and agentConfig options, see the Chat API documentation.

Known issues and limitations

• The new agentic engine only applies to Assistant.
• Data analysis
  • Is only supported on small files < 1 MB that have not been truncated.
  • Charts and graphs may not render as expected.
• Follow-up questions may appear in reasoning tokens or suggest tools that may not be enabled.