LLM is just one part of a coding agent; the harness is the real system
blaizedsouza · x · 2026-08-20
The post argues that the LLM is merely a component of a coding agent; the real system is the "harness" surrounding it.
The core loop is Action → Observation → Repeat, but making it useful requires the harness to handle:
- LLM providers
- Memory + skills
- LSP feedback
- Sandboxing + permissions
- Context compaction
- Evals + observability
This harness is exposed in two modes: an interactive TUI session and remote parallel execution. While the model decides what to do, the harness ensures those decisions work.
More from coding & agent
- Codex ports Quake to browser in 3 days using TypeScript and PlayCanvas — willeastcott · 2026-08-20
- Autonomous OS released: An 'Android for robots' — dee_hw · 2026-08-20
- Context Windows Are Not Memory: A Student Builds an Open-Source LLM Memory Framework — Cohere_Labs · 2026-08-20
- Dev Test: Replit Free Mode Handles 5-Hour Dev Session Without Hitting Limits — amasad · 2026-08-20
- safe-fix-mcp: Test-Driven Safety Gated Dead Code Remover — pamela-0 · 2026-08-20
- Grok Bot Test: Learned Expense Reporting Workflow After a Single Demo — eyishazyer · 2026-08-20