AI Systems

The useful unit is no longer just a model. It is a working system around a model:

intent and rules
      ↓
agent harness ── model
      ├── skills     reusable ways of working
      ├── MCP/tools  access to external capabilities and context
      ├── workspace  files, state, and artifacts
      └── evals      evidence that the result is good enough

Core Notes

Note	Question it answers
MCP	How does an agent gain standardized access to tools and context?
Skills	How should a repeatable workflow be packaged for an agent?
Agent harness	What actually runs the loop, enforces boundaries, and verifies outcomes?

Separation of Responsibilities

• Put stable personal and repository rules in the agent instruction layer.
• Put a repeatable, bounded method in a skill.
• Use MCP when the agent needs a standardized live connection to another system.
• Put orchestration, state, approvals, retries, tracing, and evaluation in the harness.
• Put fast-changing product comparisons in the Frontier Radar, not in these concept notes.

This separation keeps a provider change from forcing a rewrite of the whole workflow.