Skip to main content

Working with agents

You interact with agents through chat conversations. To start a new chat, select an agent from the homepage. This can be an agent you created or one shared with you.

If you already have ongoing conversations, open the Chat tab to view and manage them. On the chat page, enter a question or prompt in the message box at the bottom of the screen to receive a response from the agent. Each chat represents a single conversation thread with its own context.

You can continue a conversation as long as needed and return to previous chats at any time. This allows you to maintain context without starting over.

Chat runs on a conversation engine tuned for longer threads, tools, and streaming: answers tend to stream more smoothly, and if you refresh the page or lose connection briefly mid-reply, you are less likely to lose the whole answer than in earlier releases. If something still looks wrong, try again after a short wait—upstream model providers sometimes have incidents.

Stay oriented in long replies​

Scrolling a long answer should not leave you guessing which question produced it. Your prompt stays pinned to its turn while you scroll through the response, in a compact single-line form that does not crowd the answer. Message timestamps show a short relative time in the thread; hover a timestamp to see the full date and time.

In chat options, Thinking and Reasoning Effort settings appear only when the selected model supports them. If those controls are missing, the model simply does not expose that capability.

For optimal results, try to frame prompts that are aligned with the knowledge bases or data stores you have configured. In case you don't get the intended response, you can try new inputs and instructions. Through such engagement, agents constantly learn and add new information to their knowledge base.

Using agent capabilities during a conversation​

An agent’s responses depend on how it has been configured. Based on enabled tools and data sources, an agent may ask follow-up question, reference connected data sources, or perform actions using tools.

You can upload files directly into a chat. Agents can use these files to answer questions, summarize content, or perform analysis, depending on the tools and data access configured for the agent. Uploads stay private to your organization; if you open or download a file later, access is time-limited rather than a permanent public link. When web search is enabled for the chat or agent, answers that use the web include citations to the real source pages, with duplicate links collapsed.

From an active chat you can also access additional options such as sharing the agent with others, resetting the chat context, and deleting the chat.

Context usage in chat​

While you chat, a context usage ring may appear in the message input area. It shows how much of the active model’s context window the current conversation is using.

Open the breakdown​

Click the ring to open the Context usage popover. You see:

  • How many tokens are in use versus the model’s real context limit for this thread, with the soft max (your budget percentage of that window) called out when budgets apply.
  • A colored breakdown by message type (for example user messages, assistant replies, tool results, and attachments).
  • An Estimated label when the count is approximate rather than exact.

The ring and breakdown scale to each model’s actual window, so switching models updates what “full” means. The ring changes color as usage grows so you can spot conversations that are getting full.

Adjust your chat budget​

In regular chat (not agent-backed threads with a fixed cap), the popover can include a control to configure your personal chat context budget—the percentage of the model window reserved for history on this surface. Use it when you want more room for the current thread or prefer to leave headroom for long tool outputs.

Organization administrators can set default Chat context budget and App Builder context budget percentages in organization settings. Your personal override applies on top of those defaults for chat.

Models with very small context windows use the full window automatically; the percentage control is hidden in those cases.

How this relates to compression​

When a conversation approaches its limit, the platform may also summarize older messages automatically. Context usage shows live fill level; compression reduces what is sent to the model. Both work together so long threads stay usable.

⌘J