Key Lessons on Building Agentic Graphs: Parallelism and Feedback Loops

Nek_12 · reddit · 2026-08-29

The author shares lessons learned from 4 months of building agent graphs (workflows):

Parallelism Isn't a Silver Bullet: In cyclic graphs, parallel checks (code review, QA, architecture) cause duplicated work and increased costs. For example, fixing code review issues may require re-architecting. The solution is sequential execution based on dependencies (e.g., architecture before code review).
Agents Must Challenge Feedback: Early workflows enforced all feedback blindly, but reviewers can be wrong. Now, agents can ask the user or discuss with other agents (e.g., PM agent) to resolve conflicts, avoiding infinite loops. The author argues this is a harness issue, not just a model one.

The post also references the author's open-source orchestrator, Kent.sh.

More from coding & agent