how to choose models for each agent and subagent?

[ramonskie]

i just used the installer.

and was wondering how you can configure the model for each agent?
its all using my default model in my case copilot sonnet 4.5

[darrenhinde]

In your agent.md , you can set it by setting the model.
Like :

model: anthropic/claude-sonnet-4-20250514
temperature: 0.1

You can find the provider and model id from: https://models.dev/?search=open

From the opencode docs:

You can also define agents using markdown files. Place them in:

Global: ~/.config/opencode/agents/
Per-project: .opencode/agents/
~/.config/opencode/agents/review.md
---
description: Reviews code for quality and best practices
mode: subagent
model: anthropic/claude-sonnet-4-20250514
temperature: 0.1
tools:
  write: false
  edit: false
  bash: false
---

You are in code review mode. Focus on:

- Code quality and best practices
- Potential bugs and edge cases
- Performance implications
- Security considerations

Provide constructive feedback without making direct changes.

[ramonskie]

ahh okay. i understand now.
basicly OAC is more like a scaffold setup right?.
so you actually need to put in effort to understand and adjust it for your own needs.

this is not what i understood from the documentation. so it was a littlbit confusing.
as i was expecting a more oh my opencode approach.

[darrenhinde]

Precisely, its more the full system is exposed than hiding behind a plugin setup. It's meant to be able to be configured freely.

I need to decide if I want to go the full plugin approach for a more concentrated version but will mean not everyone will have access to the underlying model as much which from most people seems to be the case.

If people are expecting more of an oh my opencode approach, this is valuable feedback, thanks.

[ramonskie]

i just expected that behaviour because it was not documented properly.
but now that i also saw the OAC vs oh my opencode in #116 (comment)

i understand the idea more clearer now.

so i think the only thing really lacking was a clear explanation of its purpose.
and maby some extra documentation about how to configure the models and adjust the agents a littlebit to there needs
with just a few examples.
at least so that first time users would understand what the idea is.

i think your approach is much cleaner now that i had a few minutes to think about it.

[darrenhinde]

Thanks for the feedback, updated the README to be a bit more descriptive and walk through the project concepts to be more clear.

I do think will be moving towards the plugin architecture for future versions to enable certain behaviours with more accuracy but will also tend to people's expectation of experience.

[vidonnus]

@darrenhinde -- I would actually recommend setting the models in opencode.json like so (short excerpt):

  "agent": {
    "build": {
      "model": "anthropic/claude-sonnet-4-6",
      "disable": true
    },
    "plan": {
      "model": "anthropic/claude-opus-4-6",
      "disable": true
    },
    "OpenAgent": {
      "model": "anthropic/claude-sonnet-4-6"
    },
    "OpenCoder": {
      "model": "openai/gpt-5.2-codex"
    },
    "OpenCodebaseAgent": {
      "model": "openai/gpt-5.2-codex"
    },
    "TaskManager": {
      "model": "anthropic/claude-opus-4-6",
      "mode": "subagent"
    },
    "ContextScout": {
      "model": "anthropic/claude-sonnet-4-6",
      "mode": "subagent"
    },
    "ExternalScout": {
      "model": "anthropic/claude-haiku-4-5",
      "mode": "subagent"
    }
  }

This avoids conflicts on upgrade as long as agent names don't change.

[amebrahimi]

Hi, I'm using a fresh install of OpenAgentsControl, I have changed the models in my coder-agent. But when I start giving tasks to the OpenCoder it's not delegating the tasks to the coder-agent. And i'm using DeepSeekV4Pro for the OpenCoder
If it was a smarter model it would delegate the coding to the coder-agent or I'm picking something wrong here?

[canomogollon]

Hey @darrenhinde and @javi-ara — sharing our production experience with per-agent model routing in OAC (OpenCode 1.15.11 + OAC 2.1.0, 13 agents active).

The model: field in agent config works for the top-level agent, but the real power is routing different tasks to different models based on context window and capability needs. Here's what we landed on after resolving InvalidParameter errors from context overflow:

6-Tier Routing Table

Agent Role	Agents	Model	Context Window	Why
Ultra-context	ContextScout, ExternalScout, DocWriter	qwen3.6-plus	1M tokens	Massive doc extraction, runbook generation
Vision/UI	FrontendSpecialist	qwen3.5-plus	1M tokens	UI component analysis, visual review
Coding	CoderAgent, TestEngineer, DevOpsSpecialist	qwen3-coder-plus	1M tokens	Code generation with full project context
Task Planning	TaskManager	qwen3-coder-next	262K tokens	JSON task breakdown, dependency graphs
Reasoning	BuildAgent, CodeReviewer, historian, General	glm-5	200K tokens	Short-context reasoning, validation
Background	dreamer (Magic Context)	gpt-5-mini	192K tokens	Cross-session memory, async tasks

Example Agent Config

// .opencode/opencode.json
{
  "agents": {
    "ContextScout": {
      "model": "qwen3.6-plus",
      "permissions": { "*": "deny", "read": "allow", "glob": "allow", "grep": "allow", "bash": "allow" }
    },
    "CoderAgent": {
      "model": "qwen3-coder-plus",
      "permissions": { "*": "deny", "read": "allow", "write": "allow", "edit": "allow", "bash": "allow" }
    },
    "CodeReviewer": {
      "model": "glm-5",
      "permissions": { "*": "deny", "read": "allow", "glob": "allow", "grep": "allow" }
    }
  }
}

Key Lessons

1. Context window matters more than "smartness" for extraction tasks — a 1M model on a 500-file repo avoids the InvalidParameter: Range of input length should be [1, 202752] error that glm-5 hits.
2. deny-by-default permissions prevent tool hallucination — each agent only gets the tools it actually needs.
3. The orchestrator (OpenAgent) can use a reasoning model (glm-5) and delegate heavy lifting to specialists with larger context windows.

Happy to share the full 13-agent config if it helps for docs/examples. ⚡