Are AI Agents Repeating Already Solved Software Engineering Problems?

Meher_Nolan · reddit · 2026-07-08

The author reflects on whether AI agent workflows are reintroducing problems that software engineering solved long ago. Once an agent moves beyond the "Hello World" stage, its behavior depends on prompts, tool permissions, memory, retrieval settings, and currently available model endpoints. A massive amount of state is dynamically assembled at runtime and buried within framework abstractions, making review, reproduction, and auditing much harder than with static code workflows. Decades of mature software engineering practices—such as version control, CI/CD, PR reviews, rollbacks, and environment isolation—lack equivalents for agents. Agent behavior resembles dynamic runtime assembly rather than being managed as a formally versioned artifact. The author notes that GitHub Next's Agentic Workflows, gitagent, and Claude Code are already moving toward git-native workflows. While the direction is gaining momentum, the ecosystem has yet to converge, and the author asks how others are handling this in production environments.

Related event: AI Agent Workflows Face Software Engineering Test Challenges(2 posts)→

Original post →

More from coding & agent

coding & agent channel →