My Agentic Coding Workflow (June 2026)

Workflow Overview

Things That Changed

I no longer use Cursor. I'm back on VSCode (but have barely written a line of code all year).
The Codex desktop app is where I orchestrate everything.
OpenAI's GPT-5.5 is the best model (let's ignore Claude Fable for now!).
OpenCode still has the best CLI, but I rarely use it anymore.

My Current Workflow

Talk with Codex: (GPT-5.5 on High) about what I'm trying to build
Generate a Plan: Based on the conversation using the Superpowers brainstorming skill.
Sub Agents: Ask Codex to make a note in the plan about spawning a GPT-5.4-mini sub-agent for file changes
Review: Check the output, make commits
Re-Review: Ask Codex (GPT-5.5 on High) to review the output to make sure it aligns with the plan
Rinse and Repeat: If there are findings, make a new plan to fix them (repeat until no more findings)

1. Talk with Codex

I spend at least 10–20 minutes talking with Codex (GPT-5.5 on High) about what I'm trying to build and achieve.

I don't tend to use skills here, just a straightforward conversation.

One thing I've found incredibly helpful is asking Codex to push back and give me recommendations on what it thinks we should do.

2. Generate a Plan

I then stick it in plan mode (GPT-5.5 on High) and let it figure things out.

It might ask me a few questions along the way, depending on how complex the feature is.

If it's heavily UI-based or has complex functionality, I'll reach for Matt Pocock's grill-me skill to fully flesh out the finer details.

It helps a lot during implementation if Codex doesn't have to guess what you want.

3. Sub Agents

Specifically asking Codex to make a note in the plan about using sub-agents (GPT-5.4-mini on Medium) is helpful for a couple of reasons.

The first is token efficiency. I've found I use fewer tokens since the sub-agent doesn't need much reasoning, and GPT-5.4-mini has a much lower cost per million input tokens compared to GPT-5.5.

The second is context. Delegating work to sub-agents tends to keep the orchestrator agent (GPT-5.5 on High) on track as it's not getting bogged down in all the file edits.

These points will become less relevant as models get more advanced, but for now it seems to help.

4. Review

This is really the only manual part of the process.

I check the files if necessary and run the tests myself (and check in the browser if they were visual changes).

I then specifically ask Codex to "Make commits for everything you did". This makes sure it makes separate commits for each part of the plan.

5. Re-Review

This is the part where Codex does its own review. I usually ask it something like "Refer back to the original plan to make sure nothing was missed or overlooked during implementation".

It's quite common it'll come back with 2 or 3 findings.

I then ask it to make a new plan to fix those findings and implement it.

6. Rinse and Repeat

Now it's just a case of repeating step 5 until there's nothing left to find.

Sometimes you end up going down a rabbit hole of fixing non-issues forever, so I always make sure to check the findings to see if they're worth remedying.

Closing Thoughts

Things have changed dramatically since January when I posted my last update in January.

I now use a mostly different set of tools and skills and I'm iterating on changes 10x faster than I was before.

As the models get better, and my own prompting/management skills improve, this is only going to get more streamlined day by day.