Finding a Workflow After Copilot
I did not go straight from GitHub Copilot, which I wrote about in Building with Copilot, to one permanent replacement. After Copilot changed how it measured usage, I experimented with free models and different coding-agent tools. I used OpenCode with models such as DeepSeek-V4-Flash and Big Pickle, and 9Router to map models to agents, build combinations, and route work through different providers.
Those experiments showed me that the model is only one part of the experience. The workflow around it matters as much: how the agent explores a project, how it explains its plan, where I can check its assumptions, and how it verifies the result.
I sometimes used Google Antigravity too, and I liked its implementation plans and review checkpoints. Being able to inspect each stage made the agent easier to trust, and it left me with a habit I still keep: planning and implementation are separate decisions, with a human review in between.
Plan, Review, Accept, and Implement
My usual Codex workflow is simple, and each step has a purpose:
- Plan with GPT-5.6 Sol at high reasoning. The agent inspects the project, works out the request, finds the affected areas, and proposes an implementation before it changes anything.
- Review the plan myself. I check its assumptions, scope, technical direction, and edge cases, and whether it solves the ticket at all.
- Accept the plan only when it looks right. If something is incomplete or too broad, I keep refining it.
- Implement with Sol at low reasoning for routine execution, where the direction is clear and speed matters.
- Use GPT-5.6 Terra at high reasoning when it fits, depending on how complex the task is and what shape it has.
- Review and verify the result. Generated code still needs testing, judgment, and someone responsible for it.
These settings are not rules for every project. Some tasks are small enough to solve directly, and others need more exploration or another review round. Sol high for planning and Sol low for implementation is the pairing I use most, with Terra high as an alternative when the work benefits from it.
From Two Days to One or Two Hours
I can feel the difference in my working hours. A moderately complex task that used to take me around two days can sometimes be done in one or two hours. Not every two-day task becomes easy, and I am not claiming a universal productivity percentage. I see this result when the requirements are clear, the project is understandable, and the agent works from a good plan.
The saved time gives me room for urgent client requests without leaving other tickets untouched. It also leaves more attention for the decisions that matter: whether a feature fits the business need, whether the implementation is safe, and whether it will still be maintainable after the deadline has passed.
I use this approach across different stacks. Codex helps me write custom PHP scripts, work on Odoo projects where most of the code is Python, build Next.js applications, and handle broader full-stack work for my company. Switching between them is easier when the agent studies the repository first and gives me a plan I can challenge before it writes anything.
A Trial That Became Part of My Work
This phase started in August 2026 with a ChatGPT Pro trial. GoPay placed an Rp1 verification charge and refunded it, so the trial cost me nothing, and it gave me enough time to test the tools on real work instead of judging them from a few prompts.
The coding limits have felt generous so far. In a demanding session I can use around 80% of a five-hour window while the weekly meter barely moves, and I have not hit the full limit yet.
I was also given three earned rate-limit resets at unexpected moments. They came before I had used up my allowance, sometimes while the weekly meter sat somewhere around 20–70%. I don’t know why they were granted, and I don’t assume every Pro account gets them. I only know that seeing more room open up in the middle of work made me happy and grateful.
More Than a Coding Tool
Pro has been useful beyond writing code. Longer conversations are more practical because more context about a problem stays available. I use ChatGPT Work for reports, deeper research, and other tasks that need more than a quick answer. Image generation has impressed me too, especially when I need to turn an idea into a picture without breaking off from everything else.
Used together, these tools connect. A request can start as research, become a plan, turn into a software change, and end as a report or an image. I keep my attention on what the work is for, and the agent helps with the effort of moving between those stages.
The review loop is still what makes this work. I don’t accept a plan because Sol wrote it, and I don’t ship code because an agent finished quickly. I accept a plan when it makes sense and ship when the result has been reviewed and verified. Codex and the GPT-5.6 family give me leverage, and the responsibility stays with me. I later moved this workflow to a newer model, described in Coding with GPT Astra.
So far this has been a much better agent experience for the way I work. It helps me respond when demand is high, finish complex tasks in a fraction of the time, and spend more of my day on judgment than on repetitive implementation. The speed is exciting. What I value more is that I can serve my clients and support my company without the quality of the work slipping behind the urgency.