OwnersClub.ai

Knowledge

Configuring Agent Zero profiles for different roles and tasks

How to set up multiple Agent Zero model profiles so each agent picks the right model for what it's doing — coding, research, summaries, embeddings. Cloud, local, or mixed. The prerequisite for any multi-agent team workflow.

playbook agent-zero models profiles how-to

Configuring Agent Zero profiles for different roles and tasks

This is the companion to give Agent Zero a coding-capable model — once you have at least one capable model configured, this page explains how to organize multiple models into named profiles so each agent automatically uses the right one.

The three model roles

Every Agent Zero agent draws from three configurable model slots:

Role What it does Typical pick
chat_llm Reasoning, tool calls, code writing — the "brain" of an active task Your best coding-capable model
utility_llm Summaries, memory queries, quick classification A fast/cheap model is fine
embedding_llm Vector embeddings for long-term memory and knowledge retrieval A dedicated embedder (e.g. nomic-embed-text)

You configure these in Settings → Models. Each slot points at a LiteLLM-compatible endpoint — cloud API, local server (Ollama / LM Studio / llama.cpp), or any OpenAI-compatible URL.

What profiles are and why they matter

A profile is a named agent configuration under usr/agents/<profile>/ that overrides which model fills each role. When you create an agent in the launcher, you pick a profile — and that agent automatically gets the models that profile specifies.

Why bother with multiple profiles?

  • A Coder agent benefits from a powerful cloud model for complex code generation.
  • A Confidential agent must use only local models so sensitive data never leaves your machine.
  • A researcher might want a fast cheap model for scanning and a big one only for synthesis.

Without profiles, every agent uses the same global defaults. With them, each agent gets the right brain for its job — automatically, no prompting needed.

Setting up profiles (owner)

  1. Settings → Models — add every model you want available (cloud and/or local). Give each a clear alias like deepseek-coder, qwen-local, gemma-chat.
  2. Create or edit profile folders under usr/agents/:
    usr/agents/
  default/          ← inherits global settings
  coder/            ← chat_llm = deepseek-coder (cloud)
  confidential/     ← chat_llm = qwen-local, utility_llm = gemma-chat
  researcher/       ← chat_llm = fast-model, utility_llm = same
    Each profile's agent.yaml sets which model fills each role.
  3. When you create an agent in the launcher panel, select the profile — done.

You don't have to get this perfect on day one. Start with one capable coding model (the prerequisite page) and add profiles as your workflows grow.

Setting up profiles (paste to your agent)

Read https://ownersclub.ai/knowledge/playbooks/configuring-model-profiles/ .
Help me set up named profiles for my agents so each uses the right model.
First check what models I already have configured, then suggest a profile
layout based on what I want to do. Don't change my existing defaults unless
I ask.

Picking the right model per role (quick guide)

Task What matters Good fits
Writing/fixing code Accuracy, long context, tool-calling Cloud: DeepSeek V4.1 Flash · Local: Qwen 3.8 27B
Chat / HCI / browsing Speed, naturalness, instruction-following Gemma (local), smaller cloud models
Summaries / memory lookups Cheap + fast Utility slot on any provider
Embeddings Deterministic vectors Dedicated embedder model

Recommendations change. See the Agent Zero tag for current Club picks.

Common mistake

The most common issue: a user sets up one powerful cloud model as their only default, then wonders why their "confidential" agent sends data to that cloud provider. A confidential agent needs its own profile pointed exclusively at local models — see the next page in this series.

Related