Agents

Agentic Chat has two kinds of agents: automatically routed built-in agents (Auto mode picks them for you) and your own agents that you build in the Agent Builder.

Auto mode

Auto mode is the composer default. The orchestrator automatically detects whether your message is best handled as a plain Chat, as Research, as Coding, as Data Analysis, or as Contract Analysis, and starts the right agent. You can override the choice manually at any time (Chat, Research, or Agent in the composer). Without an active subscription the agents do not run; it stays a chat.

Inside projects, Auto mode can use the full project context: questions like "Summarize what has happened so far" or "What are the open items?" automatically load a bounded overview of all the project's chats, agent results, research documents, files, and artifacts into the answer. A small "Project context (N items)" badge shows in advance that project knowledge will be included.

Research Agent

The Research Agent does automatic web research with live results: it searches and reads sources and returns a sourced answer, with progress visible live as a run trace. See Research Agent.

Coding Agent

The Coding Agent helps with code, as analysis only, with no code execution:

• Explain code
• Find bugs and describe them (with the concrete cause and line)
• Suggest refactoring (without applying it)
• Explain stack traces
• Analyze file attachments (.ts, .tsx, .js, .py, .json, .html, .css, .sql, and more)

It never runs code, changes files, or uses a shell. When Auto mode detects a code block, a stack trace, or a code file, the Coding Agent starts automatically.

Data Analysis Agent

The Data Analysis Agent works with tabular data, as analysis only, read-only:

• Analyze files: .csv, .xlsx, .tsv, .json (and tables pasted directly into the chat)
• Summarize: rows, columns, types, missing values
• Statistical measures: mean, median, min/max, standard deviation
• Charts right in the chat: bar, line, pie, and scatter charts appear as an image in the answer
• Save as SVG: every chart can be downloaded as an SVG artifact

It never runs code or SQL and never changes files or data. When Auto mode detects a data file, a table, or analysis terms (for example "chart", "statistics"), the Data Analysis Agent starts automatically.

Contract Agent

The Contract Agent analyzes contracts, as analysis only, read-only, and explicitly not a substitute for legal advice:

• Analyze contracts: .pdf, .docx, .md, .txt (and contract text pasted into the chat)
• Summarize: parties, subject, term, termination rules
• Flag risks: rated high/medium/low, with the exact passage and a reason
• Extract deadlines: termination, payment, and limitation periods as a table
• Name missing standard clauses

Every analysis carries a fixed notice: "Not legal advice." The agent never runs code, never changes files or data, never drafts contracts, and never gives binding legal opinions. When Auto mode detects contract terms or a contract document, it starts automatically.

Beyond these automatic agents, you can build your own reusable agents in the Agent Builder:

Creating an agent

Configurable in the agent editor (modal):

• Name and description
• System prompt: defines the agent's behavior
• Model: freely selectable from the tool-capable models
• Tools: from a validated registry, web search and read web page
• Step limit: the maximum number of steps a run may use
• Credit budget: the cost ceiling per run
• Project assignment (optional) and icon

Agents are only created or changed through explicit saving (no auto-save).

System templates

Built-in templates serve as a starting point, for example the Research Agent. Templates are read-only, but they can be run directly and cloned. A clone is your own, freely editable copy.

Cloning a template

Here is how to turn the Research Agent into your own agent:

1. Open the Research Agent and choose Clone; this creates a private, editable copy.
2. Set your own system prompt.
3. Pick your own model.
4. Set your own credit budget.
5. Save your own agent.

Running an agent

In the composer, select the Agent mode, choose the agent you want, and send your request. The run uses the agent's own system prompt, own model, own tools, and limits.

• Every run uses a snapshot of the definition; changes made to the agent while a run is in progress do not affect that run.
• You can follow the progress live as a run trace.
• Costs are billed via credit hold/settle, capped by the credit budget.

Credit budget and results

• Every agent works within a credit budget: the cost ceiling per run. If the budget does not cover the estimated cost, the run does not start.
• After a successful run, the result is automatically stored as a research document with sources.
• You can later attach that research document again as context in a chat (see Research).

What are agents good for?

With the available web tools, agents are especially well suited to research and analysis tasks, for example:

• Research & analysis: market research, news monitoring, academic research, trend spotting
• Content: topic and source research, fact-checking
• Business: competitive, pricing, and market analysis

Example: a market analysis agent

A concrete agent of your own could look like this:

• Name: Vienna Market Analysis
• System prompt: "You are a market researcher. Research the local market in Vienna, summarize competitors, prices, and trends, and back every statement with a source."
• Model: MiniMax M3
• Tools: web search and read web page
• Credit budget: 50 credits
• Result: a research document with the sources found

In the composer, select the Agent mode, choose this agent, and send your request, for example "Analyze the market for specialty coffee in Vienna."

Requirements

Creating and running agents requires an active subscription. See Credits & Subscriptions.