Production agents need three layers: harness, loop, and graph
blaizedsouza · x · 2026-07-27
The production-agent stack is not one thing: harness, loop, and graph
The post argues that reliable production agents are built from three distinct layers:
- Harness: everything around the model — tools, state, permissions, memory, sandboxes, retries, and observability.
- Loop: the repeated cycle of work, evidence, feedback, and explicit stop conditions.
- Graph: the execution topology — which nodes run, in what order, with which branches, joins, and controlled cycles.
The core point is that failures should be diagnosed by layer: missing capability or state points to the harness, unreliable completion points to the loop, and uncontrolled branching or ordering points to the graph. The author says these terms are often mixed up, but they should be designed together once agents move out of notebooks.
More from coding & agent
- Testing Claude to Automate B2B Lead Gen and Booking — petewoodbridge · 2026-07-27
- Audited 50+ AI agent skills, and most had dangerous command-execution risks — aiz27 · 2026-07-27
- BrowserSmith automates four ChatGPT tabs to plan, write, review and run real apps — SouthernArcher9793 · 2026-07-27
- 100 hard-won tips for building a personal AI agent — blaizedsouza · 2026-07-27
- From 7-Year Stagnation to Patent Filed in 2 Weeks via AI Agent — pablooliva · 2026-07-27
- Rust side project speeds up old-work reconstruction by 4.67× and cuts tokens 94% — rudrank · 2026-07-27