Key takeaways

• A skill is a SKILL.md runbook loaded by progressive disclosure: only its description costs tokens up front; the full body loads on demand and drops afterward.
• MCP is a standard protocol for plugging external tools into an agent as named tools, letting one harness serve many domains. Use deferred schemas and stable tool ordering to protect context and caching; remember MCP tools self-guard outside the sandbox.
• Deterministic computation means pushing exact work (arithmetic, lookups, diffs) into code the model calls, not the model's head.
• The anti-hallucination contract: if a tool can fetch the truth, remove the model's ability to fabricate the answer. Design it out, do not rely on the model's restraint.
• All three reduce context cost and increase trust, which is what lets an agent do work that actually matters.

Original sources: the "Agentic Harness Architecture: Seven Pillars" breakdown, Anthropic's How Claude Code works, the Claude Code six-layers architecture breakdown, and "Inside the Agent Harness."

---