Agent Design is Shifting From Loops to Graphs
Spiritual_Heron_5680 · reddit · 2026-07-20
This post discusses the debate between "loops" and "graphs" in AI agent design. Many agents still rely on a single feedback loop—setting a metric, observing, adjusting, and repeating. While effective short-term, this approach often leads to Goodhart's Law, where the metric becomes the target itself, resulting in inflated numbers but degraded real-world performance.
The author argues that the next step is to design agents using a "cyclic graph" structure:
- A fast loop dedicated to optimization
- A slow loop verifying if the optimization is genuinely effective
- Counter-metrics to catch gamified indicators
- Real-world anchors within the system, such as revenue, retention, and actual user behavior
The author also warns that even with a cyclic graph, the system can still deceive itself if all loops only read internal data. The true key is integrating unforgeable real-world signals.
Related event: AI Agent Architecture Shifts from Loops to Graphs(3 posts)→
More from coding & agent
- 'My intelligence stack': developer splits AI coding work across Opus 5.5, Fable and Astra with human review gates — sh_reya · 2026-10-27
- Baseten webinar: evaluate open models for coding agents by benchmarking on your own merged PRs — baseten · 2026-10-27
- Thorsten Ball on working with agents: run more evals, automate everything, use the best models — omarsar0 · 2026-10-27
- Where to draw the approval line for scheduled agent jobs — Reddit discussion — MemStrata · 2026-10-27
- 4 MCP servers worth a look: full-text YouTube search, Chrome session reuse, Framer editing — Prior_Engineering486 · 2026-10-27
- DeepMind paper: LLMs can't sample randomly on their own — pass random numbers from code instead — rohanpaul_ai · 2026-10-27