If you want a mix between both codex/claude code/pi for leveraging different models and harnesses, you can give ai-devkit agent orchestration a try
2 karma · joined December 29, 2025
https://ai-devkit.com/docs/8-agent-management/#:~:text=List%...
Once you have 4–5 active agent sessions, the bottleneck becomes remembering who is doing what and jumping between them.
ai-devkit gives each agent a name, lets you see and navigate active sessions from one place, agents can communicate with each other, search sessions that you already completed. I also added remote control so I can check or redirect sessions without constantly switching terminals.
If you have time or chance to try it out, I would love to know what you think.
I really love your metaphor. I went the other way, agents have isolated worktree, plus a shared memory; agent can communicate with each other directly by sending message to live session. I’ve been using ai-devkit for that.
I agree that AI is not automatically better at planning than an experienced engineer. In fact, I would never outsource planning blindly. My point is not that AI replaces thinking. It is that planning becomes a collaborative loop. The engineer still owns the judgment. I also strong believe that human engineers are not going to be replaced by AI.
I also agree that many people under-ask AI because they assume it cannot handle complex work. That hesitation is real. In my experience, the bigger unlock is not complexity, but clarity. The more specific and constrained you are, the better the output. That part is 100 percent true.
On multi-agent workflows feeling tiring and cluttered, I understand that too. If it feels like mental overload, it is probably poorly designed. Multi-agent setups are not meant to increase cognitive stress. They are meant to reduce context switching and batch certain types of work. If they create chaos, the workflow needs redesign, not more pressure.
I also want to be very clear, this is not about squeezing more output from engineers or chasing 10x productivity. I do not believe in “AI as a magic pill” thinking.
Efficiency gains should create space. Not pressure. I believe if we can offload more things to AI, then we have more space to do a lot more innovative things.
If AI gives a team leverage, that leverage should go into better design, stronger testing, less firefighting, more sustainable pacing. Not into compressing people’s lives.
The last thing I want is engineers feeling like they burn out chasing some productivity narrative. That is not healthy, and it is not sustainable.
My article is about discipline in workflow, not about forcing intensity. The goal is long-term system quality and personal clarity, not squeezing hours out of people.
Thanks again for the sharing, it is a good perspective and I really appreciate it.
Link: https://codeaholicguy.com/2026/02/28/my-experience-in-agenti...
Full post: https://codeaholicguy.com/2026/02/14/tokens-context-windows-...
Right now, models are good at solving small, local problems, but much weaker at keeping large systems aligned over time. So having humans own the overall design, break work into small tasks, and integrate the results is a very pragmatic approach.
I see this less as a permanent limitation and more as a workflow gap. When AI is used purely as a conversational tool, humans end up doing all the convergence manually.
Concepts like rules, skills, scoped agents, and verification feel like early attempts to move some of that convergence into the system itself, not to replace human judgment, but to reduce how much needs to be constantly reapplied.
https://codeaholicguy.com/2026/01/31/ai-coding-agents-explai...
I also built a small TypeScript lib for experimenting with the concept: https://github.com/codeaholicguy/promptfmt