Agentic Chat
Agentic Chat is the conversational workspace where you run analysis, planning, and code changes with a Smart Engineering agent. It is available from Agentic → Chat in the sidebar and is the screen the application opens after sign-in.
Each conversation is handled by a single agent. You attach the context the agent should use (a repository, specific code, a skill, a recipe, a reference pack), choose a model, and the agent replies with streamed text, tool activity, and, when relevant, editable proposals or workflow steps that you approve inline.
When to use
- Ask questions about an indexed repository without leaving the application.
- Run a repeatable workflow (a recipe) against selected code.
- Ask an agent to draft or update documentation, a skill, a recipe, a reference pack, or another agent.
- Plan a change and review a proposed edit before it is applied.
Before you start
- Confirm the active organization. Agents, repositories, and other context are scoped to it.
- If you want the conversation grounded in code, make sure the repository is indexed and ready.
- Your chat history is stored in this browser only. It is not synced across devices or users.
Pick an agent
Open the chat and choose the agent for the conversation. Disabled and deprecated agents are hidden. The welcome message shows which agent is active ("How can <agent> help?").
You can start a New chat at any time or reopen a previous conversation from the chat list. Conversations can be renamed or deleted from the conversation menu.
Send a message
- Enter sends the message. Shift + Enter adds a new line.
- Type / to open commands: attach a skill, attach a recipe, add project context, or start a new chat.
- While the agent is responding you can Cancel the run. You can also Copy or Retry any response from its footer.
Add context
Open Add context to attach items to your next message:
| Context | What it does |
|---|---|
| Repository | Grounds the conversation in an indexed repository. |
| Code symbol | Includes a specific function, class, or file for discussion. |
| Skill | Gives the agent a reusable unit of expertise. |
| Recipe | Brings in a repeatable, multi-step workflow. |
| Reference pack | Adds curated project reference documents. |
In the same panel you can toggle Capabilities — the tool bundles the agent is allowed to use for the next message — and choose the Model and Thinking level (Standard or Extended).
Context usage and compaction
A context meter shows how much of the model's context window the conversation is using. When it gets large, use Compact context to summarize earlier turns so the conversation can continue without losing the thread.
Approvals
Some agent actions pause for your input before continuing. They appear inline and in the Required actions panel:
- Change proposal: the agent proposes creating or updating an asset (a documentation page, reference pack, skill, recipe, or agent). Review it, then Approve or Reject. Turn on Auto-approve changes to apply a given asset type and action (new items or edits) without being asked each time.
- Workflow approval: when a recipe runs inside the chat, approve each step to let the workflow continue.
Personalize
The Personalize entry opens the management area for the building blocks agents use: custom agents, skills, recipes, reference packs, and the history of agent and recipe executions. See Custom Agents and Recipe Engine.
Expected result
You should be able to hold a grounded conversation with an agent, attach the right context, follow streamed output and tool activity, and approve or reject any change the agent proposes before it takes effect.