I was trying to teach a co-worker some agentic stuff today, and this is totally true. Especially if they are using gimpy tools like Github Copilot.

The harness matters. None of my super-powered CLI + skills stuff worked like how it does in a more powerful harness like Codex or Claude Code.

Try telling someone to imagine it could do something amazing, if they switch to a completely different tool.

Same when sometimes you talk about AI with people and they say ChatGPT is absolutely horrible, gets everything wrong and never can keep track of more than 3-4 messages and so on.

Then it turns out they were on the free logged-out version, writing the most obtuse, confusing and contradictory prompts, and they're trying to do "real work" like this.

We're all having very different experiences with AI right now, and when people talk to each other about it, we're not nearly specifying what environment, tools, weights, etc are used, for these discussions to actually be valuable, because people are comparing apples to oranges everywhere.