Each agent lives in a
project. In the
dashboard, the New project wizard creates the project and its first agent
in three steps. The API creates the agent and its project in one request.
This guide builds an agent that runs on the SLNG orchestrator. To run the
agent in your own code on LiveKit, Pipecat, or a custom client instead, see
Improve an agent.
Dashboard
API
On the Projects page, click New project, then click Start
building on the Build an agent card.
The first step of New project. Choose Build an agent.
One request to the Managed Agents API creates the agent and its
project. The body carries the agent’s identity and its model config.
This is the complete body. The next steps show where each value comes
from.
agent.json
{ "name": "Support agent", "region": "eu-central", "language": "en", "system_prompt": "You are a support agent for Wayne Enterprises.", "greeting": "Hi, how can I help you today?", "models": { "stt": "deepgram/nova:3", "llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest", "tts": "slng/fish/tts:s2.1-pro", "tts_voice": "16cabdb7f8d240569aff36c9e480d783" }}
2
Pick a template
Templates prefill the pipeline, prompt, and voice for an industry. You can
change any of it later.
Dashboard
API
Keep From scratch to choose every piece yourself, or select a
template such as Healthcare or Insurance. Click Continue.
Start from scratch or from an industry template.
There is no template field. Set the prompt, greeting, and models a
template would prefill directly in the body. The speech model needs a
voice ID as well.Prefer models under the slng/ prefix where one exists. They run on
SLNG infrastructure in the region you pick, so each turn skips the hop
to an outside provider. The models allowed depend on region and
language. See the
model availability guide.
agent.json
{ "name": "Support agent", "region": "eu-central", "language": "en", "system_prompt": "You are a support agent for Wayne Enterprises.", "greeting": "Hi, how can I help you today?", "models": { "stt": "deepgram/nova:3", "llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest", "tts": "slng/fish/tts:s2.1-pro", "tts_voice": "16cabdb7f8d240569aff36c9e480d783" }}
3
Name the project and choose a region
Pick the region closest to your callers. Region sets the latency floor for
every turn of the conversation and filters which models you can use.
Dashboard
API
On the Configure your project step, enter a Project name, then
open the Region dropdown and pick a region.
The region selector on the Configure your project step.
The name and region are the first two fields of the body. The agent’s
name becomes the project’s name.
agent.json
{ "name": "Support agent", "region": "eu-central", "language": "en", "system_prompt": "You are a support agent for Wayne Enterprises.", "greeting": "Hi, how can I help you today?", "models": { "stt": "deepgram/nova:3", "llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest", "tts": "slng/fish/tts:s2.1-pro", "tts_voice": "16cabdb7f8d240569aff36c9e480d783" }}
4
Choose a language and create
The primary language filters models to the ones that support it and sets
the default speech to text and text to speech choices.
Dashboard
API
Open the Primary language dropdown, pick the language your callers
speak, then click Create project.
Primary language selector and the Create project button.
The wizard closes and opens the new project on its Listen section.
The highlighted line sets the language.
agent.json
{ "name": "Support agent", "region": "eu-central", "language": "en", "system_prompt": "You are a support agent for Wayne Enterprises.", "greeting": "Hi, how can I help you today?", "models": { "stt": "deepgram/nova:3", "llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest", "tts": "slng/fish/tts:s2.1-pro", "tts_voice": "16cabdb7f8d240569aff36c9e480d783" }}
The response is the full agent object, most of it defaults you did
not set. Keep the id. Every later call that changes the agent uses
it. The project appears on the Projects page under the agent’s
name. The highlighted lines are the ones you sent. Expand the block to
see the rest.
Response
{ "id": "1c5639ea-98f2-48bc-80c0-f8e41b2c3fca", "organisation_id": "18205771-aea9-4730-977d-cd05f807b0e8", "name": "Support agent", "system_prompt": "You are a support agent for Wayne Enterprises.", "greeting": "Hi, how can I help you today?", "inbound_greeting": null, "outbound_greeting": null, "language": "en", "region": "eu-central", "models": { "stt": "deepgram/nova:3", "llm": "bedrock-mantle/nvidia.nemotron-super-3-120b:latest", "tts": "slng/fish/tts:s2.1-pro", "tts_voice": "16cabdb7f8d240569aff36c9e480d783", "stt_kwargs": {}, "llm_kwargs": { "top_p": 0.95, "timeout": 30, "extra_body": { "chat_template_kwargs": { "enable_thinking": false } }, "temperature": 1, "tool_choice": "auto", "parallel_tool_calls": false, "max_completion_tokens": 6144 }, "tts_kwargs": { "latency": "balanced" }, "fallbacks": { "stt": [], "llm": [], "tts": [] }, "stt_final_timeout_s": null, "llm_first_token_timeout_s": null, "tts_first_audio_timeout_s": null, "failure_audio_enabled": true }, "models_validation_error": null, "idle_nudges": { "enabled": true, "first_nudge_delay_seconds": 15, "second_nudge_delay_seconds": 30, "hangup_delay_seconds": 15, "first_nudge_text": "Are you still there?", "second_nudge_text": "I'm still here. If you need a moment, just let me know. Otherwise, I'll have to end the call soon.", "final_hangup_text": "I'll end the call now. Feel free to call back when you're ready." }, "enable_interruptions": true, "noise_cancellation_enabled": true, "tts_cache_enabled": true, "tool_mode": "shared", "tools": [], "tool_refs": [], "mcp_refs": [], "runtime_variables": [], "livekit_deployment": "default-eu", "sip_inbound_trunk_id": null, "sip_outbound_trunk_id": null, "template_variables": {}, "created_at": "2026-09-09T15:25:34.707309Z", "updated_at": "2026-09-09T15:25:34.707309Z", "deleted_at": null}
A project with a linked agent that has a name, a region, a language, and
default models. It cannot take a call yet. The agent still needs its three
sections configured, and the project page opens on the first one.
In the dashboard, open the project and change the agent in its Listen, Think,
or Speak section. Through the API, send a PATCH to /v1/agents/{agent_id}
with only the fields you want to change. Use the id you kept from the create
response.
Top-level fields such as greeting update on their own. The models object is
the exception: a PATCH replaces it as a unit, so send the whole block,
including stt, llm, tts, and tts_voice, even when you change one value.
The Listen,
Think, and
Speak guides each show the models body for
their own fields.To replace the complete configuration in one request, send it with PUT. This
is how you change tool_refs and mcp_refs, which a PATCH rejects:
download the configuration, edit it,
and send it back. See
Attach a tool to an agent.
