My current coding efforts are still mostly in Python and Rust using Google’s Antigravity. Things in terms of model selection have changed for sure over the last few weeks. I tend to fall back to using Gemini 3.8 Flash on Medium vs. using either the new Gemini 4 Argon/Carbon models. One of the efforts I ran included a goal instruction and a couple of agents using the Carbon model and I burned 25% of my weekly tokens in just a couple of hours. Within Antigravity they introduced Orchestrator for both Carbon and Argon, but I’m nervous about how fast that will burn my weekly tokens. The Google AI Ultra plan has worked out well enough for me in terms of having enough weekly tokens to code for 20-30 hours of effort, but some swarm activities seem to burn tokens at a prodigious rate. Obviously, using more orchestration sounds like a good plan and the right direction, but it comes at a cost. Any large scale agent swarm that moves in the wrong direction seems to just destroy my token allocation for the week and is difficult to decouple from the code base.
One of the things worth mentioning that I have been learning is related to how I spend a lot of time waiting for efforts to conclude which means that the time to code something up and then test it has started delaying the next turn of the wheel. It’s like your ideas get queued. One of the ways I worked around that was having two different computers set up to run and work on the same project so I could have one executing a prolonged task and the other I can query and ask questions while keeping both in sync with Github. Planning and understanding the series of tasks that need to get done on a larger project is the hardest part at this point. You end up working on something and by the time that effort is going I’m already thinking about what is next, but you cannot just change the task and context with an effort running in Antigravity. Giving a running agent more and more layers of commands tends to create problems in terms of quality code execution.
Maybe the other thing I have learned that is worth mentioning is that I’m working from a coding posture where I have a distinct and well reasoned plan setup before any code development. I’m really finding that separating the plan development and coding execution seems to help move the process forward faster. Last night when I logged off I was sitting on three different plans that were pending coding. All of those got logged during the process of waiting for a previous plan to execute that ended up taking around 45 minutes from start to finish. The real problem is that the repo is now large enough the model is not able to really consider all the various components at the same time which creates ongoing problems. I end up having to do extensive rounds of bug hunting across the total code base. One of the interesting parts of the whole agentic coding process is that a model won’t just certify code as working good enough; it will generally keep finding bugs as long as you are willing to run bug hunting rounds.

