I am not sure if I should be proud or not.

I managed to get working 4 sub-agents Qwen3.6 35b on dual rtx 3090, I am using deepseek as orchestrator.

https://preview.redd.it/biksbgq0n81h1.png?width=783&format=png&auto=webp&s=cf8a4481c1ac439c3283925001c12841b8e6c2e7

They all working locally in parallel!!

Each subagent has a max context of 131072, which is "good" for the task it needs to work with.

Once the 4 builders the orchestrator is calling 4 local reviewers just to make sure the job was done correctly.

https://preview.redd.it/vlirjrcco81h1.png?width=778&format=png&auto=webp&s=fc92810cd1c0922a31f0576812051c4816bfb944

And after everything is passed, the reviewer sub-agent (cloud model will review the whole thing)

https://preview.redd.it/mt58vlkzo81h1.png?width=787&format=png&auto=webp&s=4c6b2715e9e695c0c7cda0fd80be2dbe0e9d6a5d

With this configuration I managed to have very low usage on APIs (I am now using deepseek for being cheap but $20 to chatgpt is also more than enough).

opencode.json (in case anyone is interested)

{ "$schema": "https://opencode.ai/config.json", "plugin": [ "@mohak34/opencode-notifier@latest" ], "permission": { "external_directory": { "/tmp/**": "allow" } }, "model": "local/qwen", "provider": { "local-ai": { "npm": "@ai-sdk/openai-compatible", "name": "Local AI", "options": { "baseURL": "http://localai.tailscale.ts.net:8080/v1" }, "models": { "qwen": { "name": "Qwen text local-ai", "tools": true, "max_input_tokens": 131072, "modalities": { "input": [ "image", "text" ], "output": [ "text" ] } } } }, "deepseek": { "npm": "@ai-sdk/openai-compatible", "name": "DeepSeek", "options": { "baseURL": "https://api.deepseek.com/v1" }, "models": { "deepseek-v4-pro": { "name": "deepseek-v4-pro", "tools": true, "modalities": { "input": [ "text" ], "output": [ "text" ] } }, "deepseek-v4-flash": { "name": "deepseek-v4-flash", "tools": true, "modalities": { "input": [ "text" ], "output": [ "text" ] } } } } }, "agent": { "orchestrator": { "mode": "primary", "model": "deepseek/deepseek-v4-pro", "temperature": 0.1, "options": { "reasoning": { "effort": "high" } }, "tools": { "write": false, "edit": false, "bash": false } }, "orchestrator_dashboards": { "mode": "primary", "model": "deepseek/deepseek-v4-pro", "temperature": 0.1, "options": { "reasoning": { "effort": "high" } }, "tools": { "write": false, "edit": false, "bash": false } }, "planner": { "mode": "subagent", "model": "deepseek/deepseek-v4-pro", "temperature": 0.1, "options": { "reasoning": { "effort": "high" } }, "tools": { "write": false, "edit": false, "bash": false } }, "builder": { "mode": "subagent", "model": "local-ai/qwen", "permission": { "edit": "allow", "bash": "allow", "webfetch": "allow" }, "temperature": 0.2, "options": { "reasoning": { "effort": "medium" } }, "tools": { "write": true, "edit": true, "bash": true } }, "visual_reviewer": { "mode": "subagent", "model": "openai/gpt-5.5", "temperature": 0.1, "options": { "reasoning": { "effort": "low" } }, "tools": { "write": false, "edit": false, "bash": false } }, "reviewer_local": { "mode": "subagent", "model": "local-ai/qwen", "permission": { "edit": "allow", "bash": "allow", "webfetch": "allow" }, "temperature": 0.1, "options": { "reasoning": { "effort": "max" } }, "tools": { "write": false, "edit": false, "bash": false } }, "reviewer": { "mode": "subagent", "model": "openai/gpt-5.5", "temperature": 0.1, "options": { "reasoning": { "effort": "medium" } }, "tools": { "write": false, "edit": false, "bash": false } } } } 

I remember the subagents had a custom prompt... but for some reason I don't have them.

BTW, opencode is great for this.