· Engineering
Beyond Vibe Coding: How Engineers Are Using Codex for Real Software Development
Most developers already use AI to write code — that part is no longer surprising. What’s changing now is something deeper: how engineers actually work with AI systems during real software development.
That was the focus of a recent AI in Action session led by Shashank at FlytBase, where he walked through how tools like Codex are changing engineering workflows beyond simple autocomplete and code generation. The session wasn’t about “AI writes code faster.” It was about a more practical question: what kind of engineering behavior do these tools encourage?
The Difference Between Fast AI and Thoughtful AI
Shashank compared the working styles of Claude and Codex in a way that instantly clicked. Claude, he explained, often feels like an enthusiastic engineer trying to move quickly and finish the task as fast as possible. Codex feels different — more disciplined, more deliberate. Instead of immediately jumping into implementation, it spends more time reading files, understanding architecture, asking design questions, reviewing structure, and planning changes carefully.
That slower behavior can initially feel frustrating. But over time, especially in larger or long-running codebases, it starts feeling much closer to actual engineering instead of rapid-fire code generation.
The Real Shift Isn’t Coding Faster
Shashank repeatedly emphasized that the biggest change isn’t “AI coding.” It’s orchestration — managing workflows, isolation, review systems, memory, parallel execution, and architectural reasoning around the model itself. That’s a completely different mindset from simple vibe coding.
Parallel AI Workflows Are Becoming the New Development Environment
Inside Codex, Shashank demonstrated how multiple isolated chats can operate independently within the same project — each chat essentially becomes its own working branch or experimental execution environment. One agent explores a redesign, another tests a feature variation, another refactors legacy code, another reviews pull requests. All in parallel, with isolation reducing conflicts while dramatically increasing development speed.
Worktrees Quietly Solve a Huge AI Development Problem
Without isolation, AI agents quickly step on each other’s changes — contexts get polluted, branches become messy, experiments break stable code. Worktrees solve this elegantly: each AI workflow gets its own isolated execution environment while still sharing the same repository foundation underneath.
AI Memory Is Becoming a Huge Unlock
Unlike simple conversation history, Codex maintains structured summaries across projects and workflows automatically — previous chats, task summaries, architectural decisions, workflow patterns, context from earlier sessions. New chats no longer start from zero, dramatically reducing repetitive prompting and onboarding friction.
Plugins Are Turning AI into an Operational Interface
Instead of manually moving between tools, Codex can directly interact with systems like Slack, GitHub, Gmail, Linear, and ClickUp. Shashank shared one workflow: a customer issue arrives through Slack, Codex analyzes it, generates an RCA, creates fixes inside a new worktree, and raises a PR automatically — end to end. That’s not just coding assistance anymore. That’s operational orchestration.
Automation Is Becoming a Feedback Loop
Every evening, Codex automatically reviews newly raised pull requests, analyzes comments, and updates repository guidelines and skill instructions based on recurring mistakes — an evolving feedback system where the workflow gradually improves over time. The goal isn’t just task automation. It’s behavioral improvement.
“The engineers gaining the most leverage from AI today are not necessarily the ones writing the biggest prompts.”
They’re the ones learning how to design better workflows around AI systems: structuring execution environments, managing context intelligently, building review loops, using isolation effectively, creating reusable memories, and orchestrating multiple agents together.