Agents Overview
An agent is a configured assistant that plans and executes tasks toward a goal. Unlike a workflow builder, an agent decides how to reach an outcome at runtime — it is not a fixed sequence of steps.
The Execution Loop
When an agent receives a task, the dynamic planning engine runs a Plan → Execute → Reflect loop:
- Plan — decompose the goal into steps and decide what each needs
- Execute — call tools, query knowledge bases, and run skills inside the sandbox
- Reflect — evaluate results, branch on conditions, and re-plan if needed
The loop continues until the goal is met or the agent reports it cannot proceed.
What Makes Up an Agent
| Field | Description |
|---|---|
| name | Display name of the agent. |
| instructions | System prompt describing the agent's role, constraints, and the outcomes it should produce. |
| execution_mode | How aggressively the agent plans and spends compute (e.g. balanced). |
| models | The model(s) backing planning and execution. |
| knowledge_bases | Knowledge bases the agent can retrieve from. |
| skills | Packaged capabilities the agent can recall. |
| tool_categories | Which categories of tools the agent may use. |
| suggested_prompts | Starter prompts shown to users. |
Instant vs. Planned Execution
For simple use cases — chat assistants or embedded AI features — an agent can run tool-enabled LLM calls instantly with no planning overhead. For complex goals, the full planning engine decomposes and orchestrates multiple steps. Start simple and scale up when needed.
Publishing & Sharing
Agents can be published to make them available to others, embedded as a chat widget on allowed domains, or driven through the Workspace API with a runtime key.