Loop, Harness, Graph: What Each Agent Layer Fixes and What It Breaks
Full-Presence7590 · reddit · 2026-11-16
A clean mental model for the three layers of agent engineering, arguing most agent bugs come from fixing a harness problem with a graph, or forcing a simple loop into a 14-node workflow.
- Loop (model + tools + exit condition): flexibility and self-correction, but unchecked it drifts off-scope, repeats broken calls, and burns money hallucinating.
- Harness (runtime + guardrails + context): sandboxes, timeouts, permission gates, truncation, compaction, checkpoints deliver real-world reliability; over-constrain and a smart model fights your scaffolding.
- Graph (state machines + deterministic edges): guarantees control and compliance (human approval, retries, branch SLAs), but every edge case becomes hand-wired brittleness.
Production rule of thumb: start with a tight Loop inside a strong Harness; add Graph edges only where failure is unacceptable — let the model reason inside the box, let code enforce the walls.
More from coding & agent
- I Shipped an App, a Site, and a CRM With Codex and Claude Code — and Learned Nothing — ThatGuyJord2020 · 2026-11-16
- Over Half of LLM Requests Go to Models Younger Than 60 Days, Vercel Gateway Data Shows — aruthesa · 2026-11-16
- Craft Anything is early alpha: no login, DRM or anti-cheat bypass — thisguyknowsai · 2026-11-16
- Open-source Craft Anything hits 34K GitHub stars in 16 days by letting Claude Code clone any app — thisguyknowsai · 2026-11-16
- InfoQ: 'Canvas for AI Agents' proposes a 4x4 framework for measuring agent work — bibryam · 2026-11-16
- Agents Added 64 Type Assertions in 3 Months to Silence TypeScript Errors Instead of Fixing Them — Puzzleheaded-Fox9821 · 2026-11-16