Teams & Agents

Teams are collections of agents that work together. Agents are the individual AI workers that execute tasks within workflow stages.

In This Section

• Agents -- Agent configuration: identity, perspective, observer mode, compaction, live monitoring, stage phases.
• Skills -- Reusable packages bundling tools, credentials, knowledge, and runtimes.
• LLM Assignment -- Team-level default LLM, agent-level overrides, and how the hierarchy resolves.

Teams

A team defines:

• Name — a descriptive label.
• LLM Configuration — which model and API key the team uses by default.

All agents in a team share the same LLM configuration unless individually overridden. See LLM Assignment for details.

Creating a Team

Inside a project, create a new team and assign an LLM configuration. Then add agents to the team.

Agents

Each agent has:

Field	Description
Name	Display name (e.g., "Lead Researcher")
Role	A description of the agent's purpose and behavior. This is the agent's system prompt — be specific about what it should do and how.
Skills	Tools and capabilities the agent can use during execution.
LLM Override	Optional — use a different model than the team default.
Context Compaction	How aggressively the agent manages its conversation history.

See Agents for the full configuration reference and field table.

Writing Effective Roles

The role field is the most important part of an agent's configuration. It tells the LLM who it is and what it should do. Good roles are:

• Specific — "You are a senior data analyst specializing in customer churn metrics" beats "You analyze data".
• Action-oriented — Describe what the agent should produce, not just what it knows.
• Scoped — Define boundaries. What should this agent focus on? What should it leave to other agents?

Skills

Skills are reusable packages of tools and credentials. When you assign a skill to an agent, it gains access to all the tools that skill provides. See Skills for details on platform skills, custom skills, and skill requirements.

See Tools for more on tool definitions and MCP servers.

Context Compaction

Long-running stages can exceed an LLM's context window. Compaction controls how the agent handles this. See Agents for the full compaction preset table with trigger thresholds, retention counts, and summarization behavior.

Mode	Behavior
aggressive	Compact early and often — saves tokens, may lose detail.
standard	Balanced approach (default).
careful	Compact only when necessary — preserves more context.
none	Never compact — use the full context window. Best for short tasks.