Developer documents shift away from MCPs and toward model-agnostic agent harnesses, skills, and context management as core leverage points

The author replaced MCPs (Model Context Protocol integrations) with direct CLI tools and stopped using Opencode, now primarily using Pi as their agent harness alongside Codex and Claude Code to evaluate different harness designs. For models, they use Claude 5.5 at high inference setting for most work, plus opus 4.6 and 4.7 for frontend tasks.

Instead of always-on MCPs that bloat session context, the author builds repo-specific and user-level skills (reusable workflows like commit formatting or pipeline debugging). New skills are created by hand-holding an agent through a workflow once, then using a meta-skill to automate the pattern—work is converted to skills once it's performed "more than a handful of times."

For code review, the author spawned a two-session pattern: the implementation session generates a summary and intent document, which is then fed into a separate review session to provide context. The author notes keeping the system prompt small (around 2K tokens in Pi versus above 20K in other harnesses) is critical to generating tokens aligned with intent.

The author purchased a Mac mini and deployed OpenClaw about 6 weeks ago but found it did not match their workflow preferences; however, it proved useful for exposing family members to production-grade models via Tailscale, and the Mac mini itself serves as a machine to SSH into and run workflows.