Skip to content
TriedDocs
Build agents

Tour of the agent editor

What each part of the agent editor does, where every setting lives, and how changes are saved and published.

The agent editor is where you build and test an agent. Open it from Agents by clicking an agent. Single prompt and multi-prompt agents share the layout on this page. Conversation flow agents use a canvas in place of the prompt column (see Conversation flow), and Custom LLM agents show a WebSocket URL panel (see Custom LLM). The right-hand settings sections are the same for every type.

The agent editor: header, prompt column, settings sections and test panel

Layout#

Area What it holds
Header Back arrow, agent name (click to rename), version badge, save status, Create and Simulation mode switch, Conductor, share, version history and Publish.
Agent Details strip Estimated Cost per minute, Latency range and prompt Tokens, plus a button that copies the agent ID.
Toolbar Model picker with the LLM settings button, voice picker, language picker.
Prompt column The prompt editor, Welcome Message and Multi-prompt states.
Settings column Collapsible sections for functions, knowledge, speech, transcription, call handling, analysis, handbook, workflows, guardrails, privacy and webhooks.
Test panel Test Audio, Test LLM, a simulated caller, tool mocks and test variables.
  • Name. Click the name to edit it. Press Enter to save or Escape to cancel. This sets agent_name.
  • Version badge. Shows Draft · Vn while you edit, or Published · Vn for a frozen version.
  • Save status. Every change saves automatically a moment after you stop typing. The header shows Saving…, Saved or Save failed.
  • Create / Simulation. Simulation swaps the editor for test cases and batch test history. See Test.
  • Conductor. Opens the AI assistant, which can propose edits to this agent that you accept or reject. See Conductor.
  • Share. Creates a Voice Orb link that lets someone talk to the agent in their browser.
  • Version history (clock icon). Opens Versions and environments. See Versions and environments.
  • Publish. Freezes the draft as a version. See Versions and environments.

Agent Details strip#

The strip estimates what one minute of a call costs and how long the agent takes to reply. It recalculates as you change the model, voice, transcription mode, knowledge bases and prompt.

  • Cost. Hover to see the breakdown: voice infrastructure, the LLM, the voice, telephony and the knowledge base when one is attached.
  • Latency. Hover to see the TTS, LLM and transcription parts. The popover also lists the extra Response wait time when Responsiveness is below 1, and the Pause Before Speaking delay for the first turn. These are not added to the range.
  • Tokens. Prompt tokens per LLM request. The minimum counts the prompt, tools and knowledge base context; the maximum adds about 2,100 tokens of conversation history.
The latency estimate with its TTS, LLM and transcription breakdown

These are estimates. Each finished call records its measured latency under Call History.

Toolbar#

Control Field Page
Model picker model or s2s_model on the Retell LLM LLM options
LLM settings (sliders icon) model_temperature, tool_call_strict_mode, model_high_priority, model_ultrafast LLM options
Voice picker voice_id, voice_model Voice
Language picker language Languages

Prompt column#

  • Prompt editor. The general_prompt. Lines starting with # are highlighted as headings and {{variables}} are highlighted. Type {{ to pick a variable. See Single prompt agents and Prompt engineering.
  • Welcome Message. Who speaks first and what the agent says first. See Welcome message.
  • Multi-prompt states. Optional states with their own prompts, transitions and functions. See Multi-prompt agents.

Settings sections#

Each section opens and closes on click. The number next to a title counts its items.

Section What it covers Page
Functions Tools the agent can call: end call, transfers, digits, HTTP, code, SMS, MCP, integration apps Functions and tools
MCP Servers MCP servers whose tools the agent can use Functions and tools
Knowledge Base Attached knowledge bases, Retrieval instruction, Chunks to retrieve, Similarity Threshold Knowledge base
Speech Settings Background sound, responsiveness, interruptions, voice speed, expressive mode, backchannels, reminders, normalisation, pronunciation Speech settings
Realtime Transcription Settings Transcription mode, vocabulary, languages, denoising, boosted keywords Speech recognition
Call Settings Voicemail, IVR, call screening, do-not-call, keypad input, silence and duration limits, ring duration Call settings
Agent Settings Timezone, transfer agent, contact memory Contact memory
Post-Call Extraction Summary, success and sentiment presets, custom fields, backfill, analysis model Analysis
Agent Handbook Personality and behaviour presets Agent Handbook
Workflows Pre-call and post-call workflows Workflows
Guardrails Blocked topics, jailbreak and scam protection Guardrails
Security & Fallback Settings Data storage, PII, signed URLs, retention, fallback voices, default dynamic variables Privacy and PII
Webhook Settings Agent webhook URL, events, timeout Webhooks
The Functions add menu with its tool types

Test panel#

Tab What it does
Test Audio A web call with your microphone. Click Run Test. Transfers are not supported in web calls.
Test LLM A text chat with the current draft. Each turn can be edited, rerun, deleted or debugged.
Simulated user An AI caller with a persona you describe talks to the agent.
Tool mocks Fixed results for named tools, so tests do not call real APIs.
Dynamic variables (the { } button) Variable values used by the test.

The test panel saves pending edits before each test, so you always test the latest draft. See Test.

Where settings are stored#

An agent's settings live in two resources:

  • The agent (Update agent): voice, language, speech, transcription, call settings, analysis, handbook, guardrails, privacy and webhooks.
  • The engine: for single and multi-prompt agents, the Retell LLM (Update Retell LLM) holds the model, prompt, welcome message, functions, states, knowledge bases, MCP servers and default dynamic variables. For flows, the conversation flow holds them.

Workflows are stored per agent through their own endpoints. The editor writes to the right resource for you. When you script changes, send each field to the resource that owns it.

Manage agents from the list#

The Agents page lists every agent with Agent Name, Agent Type, Voice, Phone and Edited by.

  • Create an Agent opens a menu of agent types. Each opens the Create agent dialog. The menu also has Generate from a description, which drafts an agent from a sentence.
  • Import creates an agent from a JSON file made by Export JSON.
  • The row menu has Duplicate, Export JSON, Move to folder and Delete.
  • Folders in the left pane group agents. Click + next to FOLDERS to add one. The folder is stored in the agent's folder field.
  • The Conductor box at the top builds or edits an agent from a description.
The agents list with folders, search and the Conductor box
Generate an agent from a description