Creating Agents

The Agent Editor is a full-page, two-column editor for configuring every aspect of an agent. There are no modal popups — everything is inline. The left column shows configuration sections; the right column shows a live test console.

Navigate to Agents in the sidebar → New Agent to open a blank Agent Editor, or click an existing agent to edit it.

Agent Editor Layout

┌─────────────────────────────────────────────────────────────────┐
│  Agent Editor — "content-writer"                    [Save] [←]  │
├──────────────────────────────┬──────────────────────────────────┤
│  LEFT: Configuration         │  RIGHT: Test Console             │
│                              │                                  │
│  1. Identity                 │  [System prompt preview]         │
│  2. Model Config             │                                  │
│  3. System Prompt            │  User: Hello, test message       │
│  4. Capabilities             │  Agent: I am a content writer... │
│  5. Skills                   │                                  │
│  6. RAG Sources              │  [Input field]        [Send]     │
│  7. MCP Servers              │                                  │
└──────────────────────────────┴──────────────────────────────────┘

Section 1: Identity

Name

The agent name is critical — it determines the expression variable used in workflows. If you name your agent content-writer, the output is accessed as {{content-writer.output}} in downstream nodes.

Rules:

Lowercase letters, numbers, and hyphens only
No spaces (use hyphens instead)
3–40 characters
Must be unique within the project

Good names: researcher, code-reviewer, seo-writer, security-analyst, summarizer

Icon

Click the icon selector to pick an emoji. This icon appears on:

The agent card in the Agents list
The Agent Node on the canvas
Bridge's agent inventory

Description

A brief description of what this agent does. Shown in hover tooltips and Bridge's context. Write one sentence that captures the agent's role: "Writes SEO-optimized blog posts from a research brief".

Section 2: Model Configuration

The flow here is key first: choose an API key, the provider is derived from it, and the model list comes from that key.

1. API Key

Select one of your project's LLM keys from the dropdown. The dropdown shows key names and their provider (e.g., "My OpenAI Key (openai)"). The selected key is what defines everything below it.

No key yet?

If the dropdown is empty, a project admin needs to add a key first in Settings → Keys & Secrets. See BYOK Keys.

2. Provider

Read-only — it's derived from the key you picked (there's no separate provider chooser). Selecting a key sets the provider automatically.

3. Model

Once a key is selected, the model dropdown is populated live from that key — for Azure that's your real deployments, for the others the provider's model list. Changing the key re-lists the models.

If live listing isn't available, you'll get a small curated list plus the option to type a model id by hand (the field links to that provider's official model-id reference). When you save the agent, TARX runs a quick validation call against the key — if the model/key pairing is invalid, the save is blocked with a clear error instead of failing later mid-workflow.

See Models and Providers for the full provider matrix and routing details.

Temperature

Controls the randomness of the model's output.

Value	Behavior	Use Case
`0.0`	Fully deterministic	Structured data extraction, classification
`0.3–0.5`	Focused, low variation	Analysis, summarization, technical writing
`0.7`	Balanced (default)	General-purpose agents
`0.9–1.2`	Creative, more variation	Brainstorming, creative writing
`1.5–2.0`	Highly random	Experimental, rarely useful

Max Tokens

The maximum number of tokens the model will generate in a single response. This is a limit, not a target — the model stops when it finishes or when it hits the limit.

Common values:

256 — Short responses, classification, yes/no
512 — Paragraph responses, summaries
1024 — Multi-paragraph responses, analysis
2048 — Long-form content, detailed reports
4096+ — Full articles, complete code files

Check model limits

Each model has a maximum context window and output limit. Setting max_tokens higher than the model's limit has no effect — the model stops at its natural limit.

Section 3: System Prompt

The system prompt is the core of your agent's behavior. It defines:

The agent's persona (who they are)
Their task (what they do)
Output format requirements (how they respond)
Constraints (what they should and shouldn't do)

The system prompt editor is a full-height text area with monospace font. You can write multi-paragraph prompts with markdown formatting (some models support markdown in system prompts).

Example: Content Writer Agent

You are an expert content writer specializing in technical SaaS content.

When given a research brief, you write a complete blog post with:
- An engaging title (include the primary keyword)
- An introduction that hooks the reader with a problem statement
- 3-5 main sections with clear H2 headers
- A conclusion with a clear call-to-action
- Natural keyword integration (do not stuff keywords)

Tone: Professional but conversational. Imagine explaining to a smart colleague.
Length: 800-1200 words.
Format: Markdown (use ## for H2, **bold** for emphasis).

For detailed guidance, see Writing Effective System Prompts.

Section 4: Capabilities

Capabilities are built-in tools that TARX executes server-side. Toggle each capability on or off:

Capability	What It Does
`web_search`	Search the web, return top results with URLs and snippets
`web_scraper`	Fetch and extract text content from a given URL

These are the only two agent-level capabilities. When you enable one, its tool definition is added to the LLM's tool list and the model decides when to call it.

Need HTTP calls or integrations?

Calling an arbitrary API inside a workflow is done with the http_request node in the workflow editor — it is a workflow node, not an agent capability. For databases and SaaS tools, attach an MCP server to the agent.

Only enable what's needed

Enabling many capabilities adds more tool definitions to the LLM's context, which increases token usage and can confuse models with decision fatigue. Enable only the capabilities the agent actually needs.

See Capabilities for detailed documentation on each capability.

Section 5: Skills

A skill is a reusable prompt preset. Assigning a skill to your agent merges the skill's system_prompt into your agent's system prompt at runtime. Skills are LLM-agnostic — they work with any provider or model.

Use the skill picker to browse available skills:

Scroll through the skill catalog (built-in skills appear first)
Filter by category tags (Writing, Analysis, Code, Security, etc.)
Click a skill to see its description and example prompts
Click Assign to attach it to this agent

To remove a skill, click the currently assigned skill and click Remove.

See Skills for more details.

Section 6: RAG Sources

RAG (Retrieval-Augmented Generation) sources let your agent query vector databases before each LLM call. The retrieved context is injected as system-level context, giving the agent access to knowledge beyond its training data.

To assign RAG sources:

Click Add RAG Source.
Select from your project's configured RAG sources.
Set the Top-K — how many chunks to retrieve from this source per call (default: 3).
Repeat for additional sources.

RAG sources must be configured first in the RAG Sources section of your project. See RAG Enrichment.

Section 7: MCP Servers

MCP (Model Context Protocol) servers expose tools to your agent at runtime. TARX connects to each configured MCP server when the agent runs, discovers available tools, and adds them to the LLM's tool list.

To assign MCP servers:

Click Add MCP Server.
Select from your project's connected MCP servers.
Repeat for additional servers.

MCP servers must be connected first in the MCP Servers section. See MCP Integration.

Saving an Agent

Click Save Agent in the top-right (or Save in the toolbar). Saving:

Validates all required fields (name, API key, model)
Persists the agent
Shows a success toast notification

The agent is now available in:

The agent picker in the canvas node inspector
The agent list page
Bridge's project inventory

Editing an Existing Agent

Click any agent in the Agents list to open the Agent Editor in edit mode. All fields are pre-populated. Make changes and click Save Agent.

Changes affect all workflows

If you change an agent's system prompt or model, all workflows that use this agent will use the new configuration on their next run. There is no version pinning. If you need a new version without breaking existing workflows, create a new agent with a different name.

Duplicating an Agent

From the Agents list:

Click the three-dot menu (⋯) next to an agent.
Click Duplicate.
The new agent is created with the name {original-name}-copy.
Open the copy and rename it.

Duplication is useful when you want to create a variant of an agent (e.g., seo-writer → seo-writer-gemini to test a different model).

Deleting an Agent

From the Agents list:

Click the three-dot menu (⋯) next to an agent.
Click Delete.
Confirm the deletion.

Workflows break on deletion

Deleting an agent that is referenced in workflow nodes will cause those workflow nodes to fail at runtime. Review your workflows before deleting agents.

Agent Configuration Examples

Minimal Agent (just LLM)

Name: summarizer
Model: gpt-4o-mini
Temperature: 0.3
Max Tokens: 256
System Prompt: "Summarize the following text in 2-3 sentences."
Capabilities: none

Research Agent

Name: researcher
Model: claude-sonnet-4-6
Temperature: 0.5
Max Tokens: 2048
System Prompt: "You are a research assistant. Given a topic, search the web for 
current information, visit relevant pages, and compile a comprehensive research 
brief with sources."
Capabilities: web_search ✅, web_scraper ✅

Classification Agent

Name: classifier
Model: gpt-4o-mini
Temperature: 0.0
Max Tokens: 256
System Prompt: "Classify the input into exactly one of: bug, feature, question, 
other. Respond with only the label."
Capabilities: none

Full-Stack Agent (all enrichment layers)

Name: enterprise-analyst
Model: claude-sonnet-4-6
Temperature: 0.3
Max Tokens: 4096
System Prompt: "You are an enterprise data analyst..."
Capabilities: web_search ✅
Skill: "Data Analyst" (built-in)
RAG Sources: company-docs (top_k: 5), policy-database (top_k: 3)
MCP Servers: salesforce-mcp, jira-mcp

Agent Editor Layout​

Section 1: Identity​

Name​

Icon​

Description​

Section 2: Model Configuration​

1. API Key​

2. Provider​

3. Model​

Temperature​

Max Tokens​

Section 3: System Prompt​

Section 4: Capabilities​

Section 5: Skills​

Section 6: RAG Sources​

Section 7: MCP Servers​

Saving an Agent​

Editing an Existing Agent​

Duplicating an Agent​

Deleting an Agent​

Agent Configuration Examples​

Minimal Agent (just LLM)​

Research Agent​

Classification Agent​

Full-Stack Agent (all enrichment layers)​