April 4, 2026

10 observations after four weeks with OpenClaw

It’s been a very interesting 24 hours in my OpenClaw experience.

Yesterday, my claw blew my mind one-shotting some amazing biz stuff for me.

Today, Anthropic broke my integration, and unbreaking it ruined my Saturday evening.

Given that, I thought I’d brain dump my current observations:

My trust level and dependence on OC have been too high for something so complicated and new. This stuff is BLEEDING EDGE and FRAGILE. (No regrets, tho... I still predict that agents are going to be even bigger than mobile, and being on the bleeding edge gives you a big advantage)
I don’t care about OpenClaw, per se. The MODALITY is what’s important to me: I want to chat on my PHONE with an intelligent agent I TRUST to make changes to systems I’ve given it access to. The implementation details are TOTALLY irrelevant to me.
The meteoric rise of OpenClaw was almost certainly contingent on early adopters freeloading on the Claude Max plan. Now that OC + Opus 4.6 will likely cost 10x as much per month, excitement will drop significantly among non-business users. (But I bet business users will still pay $1-2k per month - if they can TRUST IT).
I’m starting to see a pattern where I use the LLM creativity to brainstorm and write a skill, but I DON’T need or want any AI involved in RUNNING the skill. I want the skill to be totally deterministic - a bunch of scripts running on a schedule, just like the old days - but my claw doesn’t default to this mode; I have to tell it to operate like this.
My claw is TERRIBLE about error handling. It literally never lets me know when something failed. Which makes it hard to trust with anything important. It seems like it should be a default for it to annoy the crap out of me when something didn’t run (or even better, just fix it, repair the skill, and let me know).
OpenClaw is feeling more and more like using Linux as your desktop OS. I think this is reversable and there could be a kick ass open source agent (or tons of them), but it seems inevitable that there will be an Apple-esque “walled garden” version that is pretty constrained but “just works” for the top use cases of the average person.
The flip side of this is that there probably will be solid demand for agents that cater to folks who AREN’T average. For example, I could imagine building a harness (or at least a set of skills compatible with a popular agent platform) to address the needs of a non-normal person like an independent consultant running a solo operation.
Think of it like this... what if you could spend six months training an employee to be your dream assistant, and then sell clones of this person at zero-marginal cost?! It’d be like a staffing firm where you don’t have to pay the staff.
Related: Lots of thought-leader-y types have been using their books and blog and podcast episodes to train an AI model and create a digital clone of themselves to sell to their students as a “cheaper coach”. But I think the REAL opportunity for thought-leader-y types might be to train their ideal digital VA and sell clones of it to other thought-leader-y types. In other words, don’t replace yourself with a cheaper version. Create your dream digital assistant and sell clones of it to your competitors.
Skills will be the new apps. Once someone has a stable and useful digital assistant, being able to instantly give it a new skill will be worth LOTS of money. For example, let’s say my assistant is good at data analysis and writing software, but stinks at drafting marketing copy for a sales page, or creating ad creative for Instagram. How much would I pay to be able to drop in those skills? A LOT, assuming I trusted the creator of the skill. (e.g., If Seth Godin created an openclaw skill for drafting "permission marketing" sales pages, I’d probably pay $300 for it.)

Watch this space...

My low-level assumptions and opinions about digital assistants are evolving in real-time, but the high-level concept still looks like a total game-changer to me.

Yours,

—J