Anthropic Harness Post Spurs Setup Questions

The interesting part is not the model stack, it’s the control loop: you’re turning vague intent into a spec, code, and measurable critique. That’s the piece most DIY agent setups miss.

• –The planner is doing scope control, which matters more than raw model quality when a task runs for hours and drifts without structure.
• –The evaluator is the load-bearing role: it turns subjective output into graded criteria plus live testing, which is what makes iteration productive instead of random.
• –Local backends like llama.cpp or oMLX can work fine as infrastructure, but the real bottleneck is usually prompt design, rubric quality, and state handoff between agents.
• –OpenCode is closer to a practical orchestration layer than a model server; the harness still needs explicit contracts, file-based handoffs, and a retry/iteration policy.
• –The article reinforces a broader point for agent builders: reproducibility lives in the methods, not just in the model choice.