I switched back to GitHub Copilot from Google Antigravity last week… initially because with the Google AI Pro tier, I got stuck waiting 4+ days for the Claude models to open back up. Copilot provides Claude, or even the GPT-5.x-Codex suite, either of which can last me a month.
Then a game-changer happened: GPT released the
GPT-5.3-Codexmodel.
And after one thing led to another, I found I hadn’t even thought about switching back to Antigravity.
Antigravity wasn’t a bust, however, as I learned from it a consistent manner to set up rules and workflows, all of which I was able to carry back over to Copilot, pretty much as-is.
I had one bout with the first GPT-5.3 agent request: I talked it out with perplexity.ai, thought on it, and came to the realization that the new model was actually trying to teach me something. I got a bit flustered when it identified “gaps” in my workflow, but for which I construed more as common sense than gaps.
But, in my tree-level view, that was only one workflow.
What the agent was asking for was more about identifying gaps that would help any agent with any workflow. So what I did was, I extracted those “gaps” and embedded them directly into my workflow template. Since doing so, I’ve had 3 successful runs with the 5.3-Codex agent. I was pretty skeptical of a couple of those task requests, but it tackled each, and after which, we had some good follow-ups.
Perhaps the template helped, perhaps the new model, or maybe a bit of both. In any case, these recent workflows have been a small relief in comparison to some of the frustrations I’ve faced with these agents over the months.
In the end, in keeping with my love of sharing, I have published my agent task request workflow as a Gist, complete with a readme on how to use it. It should be applicable whether using Copilot, Antigravity, or whatever IDE or agent/model you use.
Hope it’s helpful in some sense.
Cheers!
– Keith | https://keithdc.com
#appDevelopment #webDevelopment #agenticDevelopment #agenticCoding #writtenByAHuman
