Agentron (Chat)
Agentron is the built-in chat assistant. You talk in natural language; it uses an LLM and tools to create, edit, list, and manage agents, workflows, tools, sandboxes, and more. With heap on it runs at Level 4 (recursive/self-building): the planner can create new tools and agents on the fly. With heap off it uses Level 1 (ReAct + tools). You can say things like “Create a workflow that runs my weather agent then my summarizer” and the assistant will call the right tools.
Where to find it
- Chat page: Open Chat in the sidebar (or go to
/chat) for full-screen chat with conversation history, ratings, and settings. - Elsewhere: The chat is available from the main navigation so you can ask for help or run actions from anywhere.
On the chat page you get:
- Conversation list: Previous chats; open one to continue or review.
- Ratings: Rate assistant replies (e.g. thumbs up/down) to improve future behavior.
- Settings: Customize the system prompt and choose which agents, workflows, and tools the assistant can see (context selection).
How it works
1. You send a message
You type in the chat input.
2. The assistant receives context
Your message plus context: where you are in the UI, relevant RAG chunks (if any), feedback from past runs, and the studio resources you’ve allowed (agents, workflows, tools, LLM providers).
3. The assistant can reply with tool calls
Structured format, e.g. <tool_call>{"name": "create_agent", "arguments": {...}}</tool_call>.
4. The system runs those tools
Creates an agent, updates a workflow, etc., and sends results back to the assistant.
5. The assistant continues or answers
More tool calls in the same or follow-up turns, or a plain-text answer.
The chat is not just Q&A: it can change your agents, workflows, and tools by calling the same APIs the UI uses.
How Chat routes (heap)
When the assistant uses heap mode, it runs at Level 4 (recursive/self-building): it does not call every tool in one step. A planner (one LLM call) first chooses an ordered set of specialists (e.g. general, workflow, agent, improvement) and can create new tools and agents on the fly. Each specialist has a limited set of tools; the system runs them in order (or in parallel where allowed) and merges context. So the assistant may route your message to a workflow specialist, then an agent specialist, without exposing the full tool list to a single LLM. This keeps responses focused and within token limits. You don’t need to configure this; it’s part of how the chat backend works.
See the Heap (specialists) page for an interactive view of the production specialist registry.
What the assistant can do (tools)
| Category | Tools |
|---|---|
| Agents | list_agents, get_agent, create_agent, update_agent, delete_agent |
| Workflows | list_workflows, get_workflow, create_workflow, update_workflow, add_workflow_edges |
| Tools | list_tools, get_tool, create_tool, update_tool |
| LLM | list_llm_providers |
| Runs | list_runs, get_run |
| Reminders | create_reminder, list_reminders, cancel_reminder |
| Other | create_custom_function, create_sandbox, execute_code, list_files, answer_question, explain_software |
| Remote | list_remote_servers, test_remote_connection, save_remote_server |
- “Create”, “edit”, “list”, “delete” studio resources → The assistant uses the matching tools.
- “Remind me in 20 minutes to …”, “remind me at 3pm” →
create_reminder(message +atorinMinutes). The reminder appears in the same chat when it fires. Uselist_reminders/cancel_reminderto manage. - “Fix”, “configure”, “populate” → It uses get_* first, then update_*.
- General or coding questions →
answer_question. - Questions about Agentron (e.g. “What is a workflow?”) →
explain_software.
Customization
| Option | Description |
|---|---|
| Custom system prompt | Override the default instructions. |
| Restore standard prompt | Revert to the built-in prompt. |
| Context selection | Choose which agents, workflows, and tools are included so the assistant only suggests or edits what you care about. |
| Improve from feedback | Use conversation ratings so the system can suggest prompt improvements over time. |
Reminders
You can ask the assistant to set one-shot reminders (e.g. “Remind me in 20 minutes to call John” or “Remind me at 3pm to submit the report”). The assistant uses create_reminder; when the time comes, the reminder text is posted into the same chat. See Reminders for details.
Summary
The chat is your natural-language interface to the studio: describe what you want, and the assistant uses tools to create or change agents, workflows, and tools. Use the Chat page for full-screen use, history, ratings, and settings. The run page supports improving from a run (retry with optional note so an improver can update agents or workflows).