Yesterday, a client rushed me, and I was stuck waiting for AI to change the code
At 10 p.m. Tuesday, I was still at home in Hangzhou, staring at a website redesign mockup. Sales wanted the new version sent to the client, while the designer was waiting for me to merge the pages. I was close to snapping: the problem wasn't a lack of tools, but that every step had to queue for my approval. I later realized that many small teams don't need people working harder; they need a workflow that can move in parallel.
Codex is changing the “finish one thing, then start the next” pattern
At the AI Engineer Code Summit in San Francisco, OpenAI engineers Dominik Kundel and Charlie Guo showed off a new way to use Codex: once a task is in the cloud, other tasks can run at the same time, without someone keeping a computer open and watching. A personal assistant can retain context across conversations and even take assignments by voice. Earlier cloud-based workflows left many users feeling awkward; this time, the experience finally started to feel natural.
My own take is that it turns a one-desk craft workshop into a long workbench shared by several people. I can review the pricing page while an engineer checks checkout. When building a mini-program, several review tasks can run in parallel. For a team of one to five people, what we really save isn't a few minutes of coding. It's the time spent waiting for people, chasing updates, and repeatedly explaining the background.
I wanted to try it today, so I did the math
There are two layers to the cost of recreating this setup: the direct cost of a Codex subscription and cloud-task usage, both of which vary by plan. I don't have a fixed total that applies to everyone, so I won't throw out a misleading number. In terms of time, scheduling one low-risk page review within the free allowance takes about 30 minutes. The technical barrier is being able to explain where changes are allowed and what must remain untouched; no coding knowledge is required. The first step is to open the Codex entry included with my ChatGPT plan, click “New cloud task,” and have it inspect one page only—not rewrite the entire website.
I got stuck here too: my first prompt was “make the website better,” and the pile of suggested changes was overwhelming. I changed it to “only inspect the mobile checkout flow, list the recommendations, and don't submit changes,” and then felt comfortable continuing. This tool isn't for everyone either. If I don't touch code and my budget is tight, there's no rush to try it.
Pick a setup based on how many people are on your team
If I'm just getting started: I'd run a read-only review of my own landing page, focusing on mobile layout, buttons, and forms, rather than rushing to rebuild the whole site. Getting familiar with how it responds matters more than chasing full automation.
If I already have one or two clients: I'd break each client's request into small, separate tasks—for example, checking errors first, then changing pages, and finally verifying the live result. I'd include the client's name and the parts that must not change in every task, while still personally confirming the delivery.
If I'm scaling up: I'd then consider parallel tasks, the cross-session assistant, and team permissions, delegating the repetitive checks. But I'd define the workflow, costs, and rollback plan first: review the results once a day, and keep human approval on every production change. If it feels right, I might pick the most annoying, repetitive task, run one trial, and only then decide whether it really belongs in my workflow.