Agent Teams
Overview
For highly complex, multi-step tasks, a single agent may struggle to maintain focus or lack the necessary combination of skills. Agent Teams solve this by orchestrating multiple specialized agents to work together.
In a team, one Supervisor agent decomposes a large task, delegates subtasks to specialized Worker agents, reviews their work, and compiles the final result.
Purpose
Agent Teams are designed for complex, automated workflows where you need:
- Specialization: Combining agents with different, deep skillsets (e.g., a Data Engineer agent + a Graphic Designer agent).
- Quality Control: Built-in review cycles where a supervisor checks work before execution.
- Parallel Execution: Completing multiple independent subtasks simultaneously to save time.
How Team Execution Works
When you assign a task to a Team Supervisor, the execution follows a strict 4-phase model:
Phase 1: Task Decomposition
The Supervisor analyzes your request and breaks it down into concrete, actionable subtasks. It evaluates the roster of Worker agents (their roles, skills, and bios) and assigns each subtask to the most qualified worker.
Phase 2: Worker Execution (Plan-Only)
The assigned Worker agents receive their subtasks and begin execution.
- During this phase, Workers operate in Plan-Only Mode.
- Read operations (e.g., searching data, exploring schemas) execute normally so the worker has context.
- Write/Destructive operations (e.g., creating dashboards, deleting files) are drafted as Planned Actions but are not executed yet.
Phase 3: Supervisor Review & Revision
The Supervisor reviews the output and planned actions from all Workers.
- If the work meets the requirements, the Supervisor approves it.
- If errors are found or instructions were missed, the Supervisor issues feedback and sends the subtask back to the Worker for Revision. (Teams can be configured with a maximum number of allowed revisions to prevent infinite loops).
Phase 4: Final Execution
Once the Supervisor approves all worker outputs, the Planned Actions are executed for real, and the Supervisor compiles a final, cohesive summary of what the team accomplished.
Execution Strategies
When creating a team, you define how the workers should operate:
| Strategy | Description | Best For |
|---|---|---|
| Parallel | Workers execute their subtasks simultaneously. | Independent tasks (e.g., "Search Slack AND search Jira"). Faster execution. |
| Sequential | Workers execute one after another. Each worker receives the output of the previous worker. | Dependent tasks (e.g., "Pull data from BigQuery THEN create a dashboard from that data"). |
Model Tiers
Agent capability is driven by the underlying AI model. In Vantage, models are grouped into three tiers. When building agents for a team, it's common to mix tiers (e.g., a Deep Thinking supervisor with Fast workers).
| Tier | Description | Typical Use in Teams |
|---|---|---|
| Fast | Quick responses, lower cost. | Simple data retrieval, formatting, or single-action workers. |
| Pro | Balanced capability and speed. | The standard tier for most worker agents requiring tool use. |
| Deep Thinking | Maximum reasoning power and context retention. | Best for the Supervisor agent that must decompose complex logic and review outputs. |
Agent Credits & Usage
Because Agent Teams involve multiple LLM calls across multiple agents, they consume more credits than single-agent interactions.
Base Costs:
- Agent Session: Starts a conversation
- Agent LLM Iteration: Each thinking step an agent takes
- Agent Tool Call: Each time an agent uses a tool
Tool Operation Costs: Different tools consume different credit amounts based on their complexity. For example:
- Dashboard operations (e.g., creating a dashboard, adding tiles)
- Workflow operations (e.g., running pipelines, adding nodes)
- Integration operations (e.g., querying BigQuery, searching Google Drive)
- AI operations (e.g., AI summary, forecasting)
Tip: To optimize credit usage, assign specific, limited skills to your worker agents so they don't load unnecessary tool sets, and use the Fast model tier for workers performing simple tasks.