Agent == LLM + Harness: the 5 components of agent scaffolding, explained

blaizedsouza · x · 2026-11-05

Milan Milanovic breaks down the formula Agent == LLM + Harness into five components:

Context: system prompt, instructions (e.g. CLAUDE.md), files, conversation history
Tool call format: the agreed format the LLM uses to specify tools and arguments (readfile, editfile, bash, search, ...)
Guardrails: rules and permissions enforced by the harness (always ok / never ok / require confirmation)
Runtime: code that executes tool calls and feeds results back to the LLM; can include MCP clients for external tools
Checks: hooks that run when the LLM thinks it's done, e.g. the test suite

An accompanying diagram shows the full loop from user task through Context → LLM → Tool call → Guardrails → Runtime → Checks back to the result.

More from coding & agent