Current agentic workflow

Agents team

Recently, I've been using Antigravity for my audio analysis software I'm working on, and it is great. But I also find interesting to switch between Antigravity, Cursor, VS Code and to capture new things, try new flows of agentic development, kind of trying to get comfy in slightly different ways.

So now I'm trying this switching, but with also terminal always open in split mode, so I can run Claude Code and Codex simultaneously, for different use cases, taking into account there differences, from what I noticed so far.

Claude Code on $20 Pro with Opus 4.6 model subscription is good for investigation of a problem or designing a solution, but it consumes the information more eager and hence burn the session tokens quicker.

Codex on $20 ChatGPT Plus with GPT-Codex-5.3 High model subscription looks more reasonable for long sessions. It's more focused on a particular issue, separating different aspects of several adjacent changes and commiting it separately, meaning it tracks them mentally as different subsets.

Artifacts

I find it okay for now to store all plans in plans/ folder, so I can refer an agent in a new session to what's done and retionale behind it. All the needed knowledge I store in docs/ folder, including ref-list.md if such is needed. For instance, for my audio analysis tool I spent a lot of time on investigation of Core Audio releted issues, which were not documented well, or domain-specific knowledge related to DSP, like Spectral Reassignment method for better FFT quality.

Also I find useful to make issue logs for specific topics I've been struggling too long. And it's useful to keep the back-track of what we've tried, what result we got, so I can point out the agent to avoid some falsy previously incorrect solutions.

Coordination

Right now I don't force them to work in parallel, cause I haven't come up with robust way of orchestrating them. Intuitively my mind goes towards markdown task list, but the question arises about race condition and correct way of syncing and tracking this list. I don't want to have a separate agent-orchestrator, I would like to have "pull new when ready", basically task-stealing, model for each agent.

Another aspect is a context window usage and a threshold when the agent needs to do compaction.

Another aspect is agent code review. Right now, I manually ask one agent to review local changes made by another agent. To make this process autonomous, the same, I will need to come up with some race-free tasks and stages (design, code, test, review, etc.) storage.

To be continued...