Skip to main content

Agents

Abstra Agents are AI-powered stages that use an LLM to reason step-by-step, call tools, and complete tasks autonomously. Instead of writing procedural code, you write a prompt template describing the task, and the agent figures out which tools to use and in what order.

Agents are triggered by tasks — just like tasklets — but instead of executing a fixed script, they run a ReAct loop (Reason + Act) where the LLM decides each step based on observations from previous steps.

How It Works​

  1. A task arrives at the agent stage with a payload (e.g. {"customer_name": "Alice", "request": "send invoice"})
  2. The prompt template is rendered with the task data using Jinja2
  3. The agent enters a ReAct loop:
    • Think: The LLM reasons about what to do next
    • Act: It calls one of the available tools
    • Observe: It sees the tool's result
  4. The loop continues until the agent calls finish or reaches the maximum number of steps

Creating an Agent​

Add an Agent stage to your workflow in the editor. Each agent has:

  • Prompt template: A markdown file (.md) with Jinja2 syntax describing the task
  • Permissions: Which tools and resources the agent can access (tables, files, connections, source code)
  • Max steps: Maximum number of reasoning steps before the agent stops (default: 30)

Agent Settings​

SettingDescriptionDefault
NameDisplay name for the agent in the workflowRequired
Max StepsMaximum number of ReAct loop iterations30
Input/OutputEnable workflow transitions to/from this agent—
PermissionsTools and resources the agent can accessNone

When to Use Agents​

Agents are best suited for tasks that require flexible decision-making — where the exact sequence of steps isn't known in advance:

  • Processing unstructured data (emails, documents, images)
  • Multi-step operations that depend on intermediate results
  • Tasks that involve searching, filtering, and acting on external data
  • Workflows where the path depends on the content of the task

For tasks with a fixed, predictable sequence of steps, use a Tasklet instead.