AI Agent
Use this when the task is too complex for a single prompt — when the AI needs to look things up, refine its understanding, and iterate before producing a final answer.
What it does
Runs an iterative reasoning loop with the AI model. At each iteration, the model can decide to search a connected knowledge base, refine its question, or produce the final answer. The loop continues until the model is satisfied or the maximum number of iterations is reached.
The result is a single output written to memory — but the path to that output is adaptive instead of fixed.
If you only need one prompt and one answer, use LLM Prompt instead. AI Agent is heavier and more expensive; reach for it only when you actually need the back-and-forth.
What you configure
| Field | What it controls | Required | Notes |
|---|---|---|---|
| Persona | Behavior description (tone, role, rules). | optional | Pick from the workspace's personas. |
| Model | The AI model that drives the loop. | required | Use a capable model; weak ones get stuck in loops. |
| System instruction | The brief for the AI. | recommended | Describe the task and the goal clearly. |
| LLM options | Provider-specific tuning. | optional | Defaults work for most cases. |
| Request (memory input) | The actual question or task. | required | This is what the agent is trying to answer or do. |
| Knowledge | One or more knowledge bases the agent may search. | optional | If empty, the agent works from its training data only. |
| Max iterations | Upper bound on reasoning steps. | optional | Default is a sensible value; raise it only if you see the agent giving up too early. |
| Output format | Plain text, JSON, or a JSON shape you describe. | optional | Same as LLM Prompt. |
| Result text (memory output) | The text content of the final answer. | required | What downstream steps consume. |
| Raw result (memory output) | The full response object. | optional | Includes information about iterations. |
📷 SCREENSHOT: The AI Agent configuration panel highlighting the knowledge picker and the max-iterations slider.
Example scenario
Research assistant inside a workflow. When a new customer ticket arrives, run an AI Agent step pointed at the product-knowledge base. The agent reads the ticket, searches the knowledge base for the relevant sections, refines its search if the first hits are off-topic, and produces a draft answer for the support team to review.
Recommendations
- ✅ Connect only the knowledge bases that are actually relevant. Each one adds latency and noise.
- ✅ Set a sensible max iterations. The default protects against runaway costs.
- ✅ Use this in workflows where you can spare a few seconds of latency. AI Agent is significantly slower than a plain LLM Prompt.
- ⚠️ Cost scales with iterations. Watch Metrics → Consumption the first few days.
- ⚠️ The agent's reasoning is opaque from outside. If you need to audit why an answer came out, inspect the job in the Jobs view, which records each iteration.
- ❌ Do not use AI Agent as a substitute for a chat-based Agent automation. AI Agent runs once per workflow step; an Agent automation runs an entire conversation.
What to do next
- For one-prompt-one-answer tasks: LLM Prompt.
- To build a chat-based automation: Agents.
- To prepare the knowledge bases this step searches: Knowledge.