The last few weeks have been erratic. Model providers are clearly figuring things out on-the-go and as a result we're seeing:
- Claude hitting limits before you finish your first coffee in the morning, followed by:
- GitHub pausing Copilot individual sign-ups and tightening usage limits
- Z(dotai is tightening its plans for all users upon renewal
- Codex randomly resetting limits
My weekly Claude limit was reset on Friday and I was out by Monday afternoon. So, I'm not complaining about the Codex limit resets. I'm here for it! But it's still... erratic?
None of that makes sense, and that's ok. It tells me one thing: no matter how bold the tech is, the 90% of things AROUND the tech that humans do will take their own sweet time to settle down.
We are more dependent on models today than our grandparents on their glasses. If you're reading this (not you grandpa), I presume you too are AI-pilled... Your morning starts by talking to agents and your day ends with putting an agent on an overnight task for you.
It's no news that we've been living on a token subsidy. My prediction is 2026-27 will be the year things get pricey AF for individuals because companies will focus on penetrating enterprises and subsidising them. Which means actually lower limits (not because of a bug; hey Claude!).
Non-technical users rarely care where their agents store data. So, God forbid, if Claude decides to actually take away Claude Code from the Pro plan OR reduce limits even further, your entire life may come to a halt.
So, you either reduce your dependence and "learn to live within your limits / below your means" as they say... OR prepare yourself to be model / provider agnostic.
Be model / provider agnostic
Prepare for a non-subsidized world where we work with a mix of subscriptions, OpenRouter, and local models. Well, not to live in the future, you could benefit from this during outages as well.
Finicky? Maybe. But isn't it fun preparing for such scenarios, and feeling insanely happy when they happen, and you 'Mark yourself safe' from the wave (millennials, remember Facebook?)
A few things I've done over the past few days:
- Audit your cloud projects: Ask your agent to list all your cloud projects. Ensure all your MCP servers and skills are included.
- Centralize locally: Ask it to set up a
~/workspace/where you and your agent will co-work, includingclaude.mdoragent.mdunder each directory.- Each project's Custom Instructions become an individual
claude.mdfile under each local project directory. - Put all your skills under
.agents/skillsand ask yourskill-creatorto default to this folder & create a symlink (aka alias) for each skill in the respective skill-folder (i.e..claude/skill/.codex/skill,.pi/skilletc.)
- Each project's Custom Instructions become an individual
- Switch to Cowork or Codex. That way, all future conversations & files remain on your device-workspace. If you want to get your hands dirty, use Pi coding agent or OpenCode. Both are open source coding agents which can be used as general purpose agents too, and both are model agnostic. HOWEVER, you cannot use Anthropic subscription with them. So, that's a bummer.
If you need help, reach out just talk to your agent.
A brief note about Pi
I absolutely love Pi. It's stupid simple, fast, keyboard-first, and well, model agnostic!
I started my career as a developer and that made me a keyboard-first person. So I do more tab tab space instead of drag & click. Pi's onboarding starts from keyboard shortcuts and it had me right there. Thank you, Mario!
Anyway, switching between models or thinking mode with Ctrl + P and Shift + tab, building custom features on the fly, its ability to work with stupid simple prompts, and tree (aka the rewind feature that actually works) are a few things that I've found insanely helpful.
It's a coding agent that knows itself well & can build itself as we go! Just like OpenClaw or Hermes agent!
It doesn't have built-in theme switcher. So I asked it to build it. I asked it to write stupid simple prompt to be 'general purpose agent' and it did. I said I want to work with Paper MCP server and it wrote an extension to work with it! Tweet
I've been using Gemini & GPT models in Pi and it's insanely fast. Gemini when the APIs are up that is! I can't objectively prove it but at least the perceived performance is super fast!
I wish Claude hadn't blocked other agents from using the OAuth subscription.
Anyway, check out Pi!
This is not a time-investment advice. Please read the docs carefully. :P