Work mode

In Work mode, Le Chat plans and runs multi-step work in a single conversation. The agent picks the Connectors and tools it needs, calls them in parallel when possible, streams its progress as it goes, and asks for your approval before any sensitive action.

1. Select the Work mode at the right of the chat window. The default is Fast.
2. Enter a prompt that describes the outcome you want (e.g., Triage my inbox from this morning, draft replies to anything that needs one, and send a summary to my team.)
3. Watch the agent work. You see each tool call, the rationale behind it, and progress in real time.
4. When the agent reaches a sensitive action, it pauses and asks for your approval.
5. The session continues across many turns until the task is complete.

Reach for Work mode when a task spans several tools or several steps. Common patterns:

• Cross-tool catch-up. Pull together email, messages, and calendar in a single run. Prepare for a meeting using attendee context, recent news, and talking points from your sources.
• Research and synthesis. Pull material from the web, internal docs, and connected systems, then produce a brief or report you can edit before sending.
• Triage and follow-up. Sort your inbox and draft replies, create issues in Linear from team and customer discussions, and send a summary to your team on Slack.
• Recurring deliverables. Build a weekly project status report from docs, trackers, and recent discussions. Produce competitive analyses, launch packages, or executive summaries.
• Data and exception handling. Process invoices or records, flag anomalies, and prepare an exception report.

Work mode runs Mistral Medium 3.5 inside an agent harness. The harness is the software loop that gives the model access to tools, manages its working memory, and runs steps until the task is complete. A few things change compared with a regular chat:

• The agent picks its own tools. You don't select web search, Connectors, or Libraries up front. The agent looks at what is connected and uses the right tools for each step on its own.
• Parallel tool calls. When steps don't depend on each other, the agent runs them in parallel, so tasks that pull from many sources finish faster.
• Multi-step plans. Improved context management, intermediate reasoning, and retry logic let the agent work through long tasks without losing the thread.
• Streaming progress. You see each tool call and the agent's reasoning as they happen.
• Canvas as the output surface. The agent writes briefs, reports, tables, and slide outlines to Canvas, where you can preview and edit them.
• Approval for sensitive actions. The agent asks for your explicit confirmation before any destructive action: sending an email, posting a Slack message, or creating or deleting a record. The exact list of approval-gated actions follows what you configured at the Connector level.

In the preview, the agent has access to the following tools:

• Connectors, including third-party Model Context Protocol (MCP) servers
• Web search
• Canvas for previewing and editing the result
• Libraries and file upload
• A bash sandbox for code execution
• Your custom instructions and preferences

• Performance. Short queries can take longer in Work mode than in Fast or Think.
• False tool-access errors. The agent occasionally reports a tool as unavailable when it is not. Retry the step.
• Email signature. Drafted emails are sometimes signed as "Le Chat" rather than with your name. Edit the signature before sending.
• Mode is locked after the chat starts. You cannot switch a Work mode conversation back to Fast or Think. Start a new chat to switch modes.

• Describe the outcome, not only the question. "Send a Slack summary to my team" gives the agent a clear endpoint; "summarize this" does not.
• Connect the tools you want it to use. Work mode is most useful when the relevant Connectors are already linked, such as email, calendar, Notion, Linear, Slack, and GitHub. Manage them under Intelligence → Connectors.
• Review approval prompts carefully. Sensitive actions only happen with your confirmation. Read the proposed action before approving.

• Think mode: step-by-step reasoning for a single complex question.
• Deep Research: long-form research reports with citations.
• Vibe Code Workflow: agentic coding on a GitHub repository in a cloud sandbox.
• Connectors: manage the integrations Work mode draws from.
• Canvas: edit the briefs, reports, and slide outlines Work mode produces.