1. Ema v2 Docs
  2. Agent Reference
  3. Respond

Respond Agent

The Respond agent (respond) generates a natural-language reply using its system prompt and LLM configuration — nothing more. It is the simplest LLM-based agent type: no search loop, no schema, no tools by default. You give it Instructions and a model, it produces text. Use it whenever a workflow step just needs the model to say something — a confirmation message, a summary, a rewrite, a friendly reply.

In the AI Employee (AIE) builder, the Respond Agent appears in the Generation group of the agent library.

First-class, lifecycle-eligible type. Unlike the Document Writer (which is dispatch-only), respond is one of the platform's lifecycle-eligible base types. It is seeded in the agent-type catalog with a live v0 version and is subject to publish-time version validation, exactly like search_respond, intent_classification, extraction, and rule_validation. See the Agent Reference overview.

When to use it

Use a Respond agent for a pure generation step where the model's only job is to write a reply from your Instructions:

  • Acknowledge or confirm something to the user ("Your request has been logged…").
  • Summarize or rephrase upstream content.
  • Produce a short message that a later node delivers.

If the step needs to ground the answer in your documents, use Knowledge Search or Search and Respond. If it needs to call integrations or return structured JSON, use a Custom agent. If it produces a full document, use the Document Writer.

How it works

respond uses the platform's passthrough injector: no type-specific instruction block is appended to your system prompt. What you write in Instructions is what the model receives (after {{input.field}} placeholders are substituted and wrapped in a prompt-injection guard). The agent then runs a single LLM completion through EmaFusion™ and returns the text.

Because there is no search step and no type-specific contract, a Respond agent is the lightest LLM type on the shared execution model — it validates inputs, resolves Instructions, calls the model, and returns the reply.

Configuration

The Respond agent exposes these advanced settings in the builder:

SettingPurpose
InstructionsThe system prompt the model follows. Supports {{input.<field>}} placeholders.
LLM configurationThe model and generation settings, served through EmaFusion™.
HITLOptional human-in-the-loop, letting the agent pause to ask a clarifying question via ask_human.

The respond base type has no required type_config — its type_config schema is empty. The default Instructions are:

You are a helpful assistant. Respond to the user based on the system prompt
and LLM configuration.

User instructions: {{input.instructions}}

Inputs and output

The preset declares one input field, instructions — the query or prompt the agent should respond to. You wire it on the node through input_mapping.

Input

{ "instructions": "Thank the customer for their patience and confirm their refund of $49 is being processed." }

Output

The Respond agent emits a bare string — the model's reply, with no JSON wrapper. There is no { "message": "..." } envelope at runtime.

Thank you for your patience! Your refund of $49 is being processed and should
appear on your statement within 5–7 business days.

Wiring the output downstream. Because the output is the whole string, reference it as {{<node_id>.output}} in a downstream node's input_mapping — without a trailing field name. In a publish node, omit field_path so the publisher takes the full string. In the builder, the node also exposes a single answer output port that maps to this whole-output value.

Example

A simple acknowledgment step at the end of a workflow:

  1. An Intent Classification node routes a request.
  2. A Respond node generates a tailored confirmation message, with the request details wired into its instructions input.
  3. A publish node sends the reply to the user.

What's next

Last updated: Jul 3, 2026