Agent Build: Configure How an AI Agent Works

Agent Build is the internal configuration surface for one AI Workforce agent. It shows how input reaches the main agent, which sub-agents it can delegate to, and which tools and skills each agent can use. Use Agent Build to change behavior. Use Agent Space to see the agent working across the Project, and use Playground to run and test it.

​

Agent Build Modes

The top navigation separates three jobs:

• Build: configure the agent graph and internal behavior.
• Run: test the draft or published agent in Playground.
• Deploy: publish or connect the agent to supported execution surfaces.

The status near the top shows whether you are viewing the Draft or published version. Draft changes do not automatically replace the published configuration.

​

Build the Agent Graph

An Agent Build can contain:

• Input: the message or instruction entering from the Run experience or a deployed surface.
• Main agent: owns the user request, chooses tools, delegates work, and returns the final response.
• Sub-agents: specialists that receive bounded tasks from the main agent.
• Tools: capabilities such as File Search, Web Search, app connectors, API requests, structured output, and custom MCP tools.
• Skills: reusable task instructions and supporting resources attached to an agent.
• Sticky notes: documentation and setup reminders placed directly on the canvas.
• Connections: define which tools or sub-agents the main agent can call.

Keep the main agent responsible for the overall outcome. Add a sub-agent when a responsibility is distinct enough to deserve its own instructions, model, tools, and trace.

​

Configure an Agent

Select an agent node to open the configuration panel. You can configure:

• Name and icon: identify the agent in Agent Space, Playground, and traces.
• Instructions: define identity, responsibilities, process, tool rules, escalation, and expected output.
• Generate with AI: draft or improve instructions, which you should review before saving.
• Model: choose a model whose capabilities match the work.
• Tools: attach only the actions and data sources the agent needs.
• Skills: upload, browse, or generate reusable task guidance.

Save changes before leaving the editor. Test one behavior change at a time so you can understand its effect in the trace.

​

Main Agents and Sub-Agents

The main agent talks to the user. A sub-agent works behind the main agent and returns its result to the main agent. Sub-agents cannot contain their own sub-agents. Use them for focused work such as research, classification, extraction, content drafting, validation, or escalation. Instructions should tell the main agent:

• What result it owns
• Which tasks it may delegate
• When it should use each tool
• When it must request human input
• What it should do when evidence is missing

See Multi-Agent Orchestration and Handoffs for delegation and routing patterns. To emit schema-validated JSON, add the dedicated Structured Output tool.

​

Draft, Test, and Publish

Use this improvement loop:

1. Change one instruction, model, tool, skill, or connection.
2. Save the draft.
3. Open Run and use a realistic test prompt.
4. Inspect tool calls, sub-agent delegation, files, human-input requests, and the final response.
5. Open the trace when the result is unexpected.
6. Publish only after the draft behaves as intended.

​

Capability Facts

• Product surface: Agent Build
• Purpose: define the internal behavior and capabilities of one agent
• Contains: input, main agent, optional sub-agents, tools, skills, notes, and connections
• Main agent: owns the conversation and final answer
• Sub-agent: receives delegated work and does not talk directly to the user
• Draft behavior: editable and testable without automatically replacing the published version
• Run surface: Playground
• Operational surface: Agent Space

​

Troubleshooting

​

The agent does not use an attached tool

Confirm that the tool is connected to the correct agent and that the instructions explain when to use it. Test with a prompt that clearly requires the tool, then inspect the trace.

​

A sub-agent is never called

Confirm that the sub-agent is connected to the main agent and has a distinct role. Tell the main agent which tasks should be delegated instead of duplicating the same responsibility in both agents.

​

The Run result does not include the latest changes

Confirm whether Playground is testing the Draft or published version. Save the Agent Build and start a new test conversation when you need a clean comparison.