The model is the least interesting part. What determines whether an agent is a collaborator or a slot machine is the harness around it — what it can see, what it can touch, how it knows it’s finished, and what survives when the session ends.
Two machines, two jobs
I run a split setup and give each half the work it’s actually good at.
| Machine | Role | Why |
|---|---|---|
| RTX 3090, 24GB VRAM | Actor — code generation, diffusion, throughput | Sprinter. Fast on quantized models, does the volume work |
| M2 Max, 64GB unified | Planner — long context, reasoning chains | Professor. Slower per token, but holds the whole problem |
Benchmarking my own runs changed what I use. A mixture-of-experts model hit 112 tok/s on the 3090 and was the obvious pick on paper. The 27B dense model runs at 35 tok/s — and holds coherence across a 262K context with no observable drift. In multi-step agent chains, accuracy saves more wall-clock time than speed does, every time. Distilled models are worse still: they compress the reasoning chain, which is fine for simple tasks and quietly disastrous on real work.
Loops that converge
The self-looping pattern — feed the task back until the agent declares completion or hits a ceiling — works beautifully on a specific class of problem: refactors, test coverage, lint, greenfield builds. What they share is a verifiable outcome. Did the build pass? Did the tests go green?
It fails on anything vague. A loop without a completion criterion doesn’t converge, it wanders, and it bills you for the walk. Most of the skill is recognizing which task you actually have before starting the loop.
Skills as memory
An agent that re-derives your conventions every session is an agent you’re paying to have amnesia. So the conventions get written down as skills and context files — narrow, composable, versioned — and the loop becomes: propose, run, evaluate, keep or revert.
Kept skills compound. Reverted ones cost one iteration. Over months, the harness becomes the actual asset, and swapping the model underneath it is a Tuesday.
Where it points
This is the same instinct behind everything else I build: don’t perfect one output, build the system that makes a hundred possible and then bring taste to the choosing. Process is the product.
Stack
Claude Code, Cursor, MCP servers, Agent Skills, Ollama and LM Studio for local serving, Qwen and other open weights, and a lot of markdown that turned out to matter more than the code.