I'm having mild success with how llamacpp works with thinking models. It's implemented `reasoning-budget` and `reasoning-budget-message`; I've paired the message with the agent, along with having a specific name for the agent. Not just "you're a researcher" but "your name is <ai agent>".
Using opencode then, this is the primary agent, and I've added dynamic context pruning plugin; so when the Qwen3.6-35B-A3B model starts getting confused as evidenced by the length of the reasoning-budget cut off, it injects a message like `<ai agent>, you've been thinking too much, compress the context and use a subagent to figure out the quandry`. So that gets injected every time the budget exceeds.
About half the time it works in either of the options, which is compressing/pruning to reduce the poisoned context or forming a specific subagent description of the problem and letting the subagent, hopefullly, work on a smalled scoped issue.
I think the next step is some kind of omniscient markdown file/folder combo which can be used to survive some of the microcosmic issues with the dumber MoE slices.
> A decent model with a great harness beats a great model with a bad harness.
Agreed. I don't use off the shelf skills and such. I craft my own that best serve the abstract problem rather than act as generic code tools.
If I am working on a game I don't want a bunch of generic language tools and skills, I want game engine specific, Vulkan specific constraints, GPU pipeline specific context that are whittled down themselves to fit the entity data model of the specific game
Gemma 4 and Qwen3.6 then crank out banger results after Gemini plans and architecture are handed to them (I scrape the AI Studio log from Gdrive rather than use the Gemini APIs)
I quit using RAG and MCP build out the harness without them.
Imo RAG and MCP were same old software industry effort of conjuring endless general purpose software things to discuss at the expense of doing a good job within reasonable constraints
I'm having mild success with how llamacpp works with thinking models. It's implemented `reasoning-budget` and `reasoning-budget-message`; I've paired the message with the agent, along with having a specific name for the agent. Not just "you're a researcher" but "your name is <ai agent>".
Using opencode then, this is the primary agent, and I've added dynamic context pruning plugin; so when the Qwen3.6-35B-A3B model starts getting confused as evidenced by the length of the reasoning-budget cut off, it injects a message like `<ai agent>, you've been thinking too much, compress the context and use a subagent to figure out the quandry`. So that gets injected every time the budget exceeds.
About half the time it works in either of the options, which is compressing/pruning to reduce the poisoned context or forming a specific subagent description of the problem and letting the subagent, hopefullly, work on a smalled scoped issue.
I think the next step is some kind of omniscient markdown file/folder combo which can be used to survive some of the microcosmic issues with the dumber MoE slices.
> A decent model with a great harness beats a great model with a bad harness.
Agreed. I don't use off the shelf skills and such. I craft my own that best serve the abstract problem rather than act as generic code tools.
If I am working on a game I don't want a bunch of generic language tools and skills, I want game engine specific, Vulkan specific constraints, GPU pipeline specific context that are whittled down themselves to fit the entity data model of the specific game
Gemma 4 and Qwen3.6 then crank out banger results after Gemini plans and architecture are handed to them (I scrape the AI Studio log from Gdrive rather than use the Gemini APIs)
I quit using RAG and MCP build out the harness without them.
Imo RAG and MCP were same old software industry effort of conjuring endless general purpose software things to discuss at the expense of doing a good job within reasonable constraints