AI Agent Overview
An AI Agent is an LLM-driven agent. Instead of a fixed graph, you give it an identity, a set of skills, tools, memory, and policies — and the LLM decides what to do at run time. It reasons about each request, picks the right tools, calls them, observes the results, and responds.
Use an AI Agent when the problem is open-ended — support assistants, research, question answering over your data, or any task where the exact steps aren’t known in advance. When you need a deterministic, auditable pipeline instead, build a Task Agent.
How an AI Agent differs from a Task Agent
| AI Agent | Task Agent | |
|---|---|---|
| Engine | LLM planning + tool calling | Visual graph of nodes |
| Builder | Form-based editor with four tabs | Drag-and-drop canvas |
| Control flow | The LLM decides the path | You wire the path explicitly |
| Best for | Chat, assistants, open-ended tasks | Predictable, step-by-step flows |
A Task Agent can be attached to an AI Agent as a tool. This lets an AI Agent call a deterministic workflow when it needs one — see Attach tools.
What makes up an AI Agent
You configure an AI Agent across four tabs in the editor. Each maps to one part of the agent’s definition.
Identity
The agent’s name, an Administrative Summary (what it does), a Persona (personality, tone, and safety constraints), the Skills it can perform, and a Fallback message for requests it can’t handle. See Agent Identity.
Skills
A skill is a named capability — a description, matching tags, a system-prompt overlay, and a set of tools the agent may use for that skill. When a user message arrives, the agent matches it to the best skill and uses that skill’s tools and instructions. Skills are the richest part of an AI Agent. See Skills.
LLM Settings
Controls how the agent reasons: the reasoning mode, step and tool-call limits, extended thinking, and how much of its reasoning end users see. You can also pin a specific model and attach an LLM policy for guardrails. See LLM Settings.
Two reasoning modes are available:
| Mode | How it works |
|---|---|
| ReAct | The agent thinks and acts in interleaved steps — good for dynamic tasks. |
| Planner → Executor | The agent plans all steps first, then executes them — good for complex workflows. |
Memory
Optional, opt-in memory across three tiers: Profile (user preferences), Episodic (past conversation summaries), and Semantic (a knowledge base plus self-learning). See Memory.
Modes
A mode is how the agent is used. You enable modes in the editor and at deployment time.
| Mode | Description | Where it’s enabled |
|---|---|---|
| Conversation | Chat sessions with memory, via text | Editor |
| Voice Starter | Real-time voice using STT/TTS or a native realtime model | Editor |
| Scheduled | Cron-triggered autonomous runs | On a deployment |
| API | Stateless REST endpoint for integrations | On a deployment |
Conversation and Voice are configured in the agent editor. Scheduled and API modes become available once the agent is deployed (added to a deployment, or deployed on its own). See Deploy.
The AI Agent lifecycle
- Create — Name the agent and pick its initial modes.
- Configure — Set identity, skills, LLM settings, memory, and modes.
- Test — Chat with the agent in the live preview; test governance with role impersonation.
- Version — Save your work as a draft version, and create new versions as you iterate.
- Publish — Promote a version to production.
- Deploy — Make it reachable at a URL, and enable Scheduled/API modes.
What’s next
- Create an AI Agent → — Name it and choose modes
- Editor tour → — The header, four tabs, and live preview
- Skills → — The core of what your agent can do