Understand task agent nodes in workflows

Task agent nodes add large language model (LLM) reasoning to IT operations automated workflows. You can place a task agent node at any point in a workflow to handle operational tasks that require interpretation, analysis, or context-aware decision-making.

When a workflow reaches a task agent step, the system sends a configurable prompt to an external LLM, which returns a response. You configure the prompt to include workflow context and data. The LLM can analyze unstructured data, produce structured output, and call tools to interact with external systems. Other steps in the workflow use the same expression and variable system to pass data into and out of the LLM.

Task agent nodes integrate with external LLMs to bring AI reasoning into IT operations workflows. Automation orchestrator does not host or serve AI models and does not replace agent frameworks or AI platforms.

NOTE: You can add human oversight to the workflow using an approval step after the task agent step.

Why use task agent nodes

IT operations automation often involves unstructured data such as error logs, support tickets, or alerts that require interpretation rather than fixed rules. Task agent nodes let you bring AI reasoning into these workflows while maintaining governance through automation orchestrator's approval gates and audit controls.

Use task agent nodes when a workflow includes:

  • Unstructured data analysis, such as interpreting error logs, support tickets, or monitoring alerts
  • Configuration drift analysis, such as comparing current system state against a known baseline and recommending corrective actions
  • Patch impact assessment, such as evaluating which systems are affected by a security advisory
  • Dynamic decision-making based on operational context that you cannot reduce to a fixed set of rules, such as selecting a remediation path based on current system state
  • Classification and routing, such as categorizing incidents by severity or type
  • Incident triage that requires analyzing multiple signals to recommend a remediation action

How task agent nodes work

  1. An automation designer adds a task agent node to a workflow and configures it with a natural language prompt that describes the task.
  2. The automation designer selects a model from an enabled LLM provider integration, then configures the LLM provider credential for that integration.
  3. Optionally, the automation designer defines a response schema to enforce structured JSON output, uploads context files for reference data, and selects which tools the LLM can access.
  4. When a running workflow reaches the task agent step, the workflow engine resolves any template expressions in the prompt. It then sends the configuration to the LLM orchestrator service. The engine resolves the model's base URL and API key from the selected integration and credential.
  5. The LLM processes the prompt. If the designer uploaded context files, the LLM reads them first. If tools are available, the LLM can call them to gather additional information from external systems.
  6. The LLM returns its response. If the designer defined a response schema, the platform validates the response against the schema. The output becomes available to downstream steps through the expression and variable system.

The system resolves all template expressions before it sends the prompt to the LLM. The LLM never receives raw expression syntax or secret references.

Governance options

Automation orchestrator validates and constrains prompts to prevents malformed or excessively large prompts from reaching the large language model (LLM). The platform enforces a 10,000-token limit on prompts. You control prompt content in the task agent step configuration using the prompt text, system message, and optional attached files that you supply.

To add governance beyond the platform limit, you can place a task agent step before the primary step to validate or filter inputs. Alternatively, place a step after it to evaluate outputs.

For example, place a task agent step:

  • Before your main analysis step to detect and redact sensitive information, such as credentials or personally identifiable information, from support tickets.
  • After a remediation step to verify that proposed actions comply with your security policies before running them.

This layered approach lets you define custom governance controls that meet your organization's requirements.

Task agent step context

When a task agent step runs, the LLM has access to the following information:

  • The natural language prompt that you configured, with all template expressions resolved to their runtime values
  • Data from earlier steps, injected into the prompt through expressions
  • Up to 10 context files that you uploaded during step configuration (PDF, DOC, DOCX, TXT, or MD format)
    Note:

    To attach files, you must configure the optional S3 storage during installation.

  • Tools that you enabled for the step, based on the selected tool access strategy: all tools, no tools, or a specific subset of tools. Tools are discovered from enabled MCP server integrations.