So my harness distributes sub-tasks to agents within 64k context, with many elements to minimize context use for fluff. The models don’t have a benchmark beyond that of “do they respond through 64K, does the response error?” If they fail at the harness subtask I don’t use them.
Long context tasks can always be divided, LLM’s are great at distributing and converging those segments. These models do not get long form multi hour tasks. The harness assembles the plan, task and sub-tasks with one-shot calls to various models, and the “swarm”(it’s sequenced on 16GB VRAM) handles the grunt work.
I developed a spec based workflow in march that I’ve used successfully with this workflow, but we built the harness greenfield, and I’ve not even kept the starting models we began with, as newer models became more capable.
Current trio:
fv78pd/gemma-4-E4B-Agentic-Opus-Reasoning-GeminiCLI-GGUF Q 4